{"items":[{"id":"claude-sonnet-5.5","name":"Claude Sonnet 5.5","providerId":"anthropic","family":"Claude 5","kind":"llm","releaseDate":"2026-09-28","knowledgeCutoff":"2026-06-01","status":"ga","summary":"Anthropic’s mid-tier Claude for everyday building, debugging and documents.","description":"Claude Sonnet 5.5 replaces Claude Sonnet 5 as Anthropic’s general-purpose workhorse. Anthropic aims it at well-defined day-to-day tasks: shipping features, fixing bugs and drafting polished documents. It takes text, images and files.","useCases":["coding","agents","writing","chat","long-context","enterprise"],"strengths":["Everyday coding","Agentic coding","Document drafting"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1000000,"maxOutputTokens":128000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":2,"output":10,"cachedInput":0.1,"batchInput":1,"batchOutput":5},"usage":{"tokenShare":1.32,"usageRank":17,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":56.2,"codingScore":48.5,"agentScore":43.8,"mathScore":38.2,"toolCallingScore":36.2,"visionScore":39.8,"sweBenchPro":81.3,"gpqa":92.7,"tauBenchAirline":78.7,"arenaText":1477,"arenaWebdev":1770,"eci":166.7},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":97,"links":{"website":"https://www.anthropic.com/claude","docs":"https://docs.anthropic.com"},"sources":["https://platform.claude.com/docs/en/about-claude/pricing","https://platform.claude.com/docs/en/about-claude/models/overview","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Anthropic","blendedPrice":4},{"id":"claude-opus-5.5","name":"Claude Opus 5.5","providerId":"anthropic","family":"Claude 5","kind":"llm","releaseDate":"2026-09-22","knowledgeCutoff":"2026-06-01","status":"ga","summary":"Anthropic’s top Claude model for hard reasoning, large codebases and long agent runs.","description":"Claude Opus 5.5 succeeds Claude Opus 5 as Anthropic’s most capable model. Anthropic positions it for difficult reasoning, multi-step changes across large codebases and agents that work autonomously for long stretches. It takes text, images and files.","useCases":["coding","agents","reasoning","research","long-context","enterprise"],"strengths":["Large-codebase refactors","Long-running agents","Text, image and file input"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1000000,"maxOutputTokens":128000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":4,"output":20,"cachedInput":0.2,"batchInput":2,"batchOutput":10},"usage":{"tokenShare":3.45,"usageRank":9,"asOf":"2026-10-31","topTasks":[{"label":"File I/O","share":2.4},{"label":"Workflow Execution","share":2.2},{"label":"Code Generation","share":2.2}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":58.8,"codingScore":50.9,"agentScore":46.1,"mathScore":42.1,"toolCallingScore":35.1,"visionScore":42.1,"sweBenchPro":89.9,"gpqa":90.2,"tauBenchAirline":75.2,"arenaText":1509,"arenaWebdev":1810,"eci":168.4},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":92,"links":{"website":"https://www.anthropic.com/claude","docs":"https://docs.anthropic.com"},"sources":["https://platform.claude.com/docs/en/about-claude/pricing","https://platform.claude.com/docs/en/about-claude/models/overview","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Anthropic","blendedPrice":8},{"id":"gpt-6-astra","name":"GPT-6 Astra","providerId":"openai","family":"GPT-6","kind":"llm","releaseDate":"2026-09-04","knowledgeCutoff":"2026-04-30","status":"ga","summary":"OpenAI’s flagship for deep research, science and end-to-end engineering.","description":"GPT-6 Astra sits at the top of OpenAI’s GPT-6 line. OpenAI pitches it for long, demanding jobs: in-depth analysis, software engineering, research, scientific work and producing finished documents.","useCases":["reasoning","math","research","coding","agents","enterprise"],"strengths":["Scientific reasoning","Deep research","Long-horizon tasks"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1050000,"maxOutputTokens":128000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":10,"output":50,"cachedInput":1,"note":"Prompts over 272K input tokens: $20 input / $75 output per 1M tokens"},"usage":{"tokenShare":0.35,"usageRank":35,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":56.1,"codingScore":46.5,"agentScore":44.7,"mathScore":33.8,"toolCallingScore":36.3,"visionScore":41.2,"gpqa":94.4,"tauBenchAirline":68,"arenaText":1476,"arenaWebdev":1773,"eci":169.1},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":88,"links":{"website":"https://openai.com","docs":"https://platform.openai.com/docs"},"sources":["https://developers.openai.com/api/docs/pricing","https://developers.openai.com/api/docs/models/gpt-6-astra","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"OpenAI","blendedPrice":20},{"id":"gpt-6.1-sol","name":"GPT-6.1 Sol","providerId":"openai","family":"GPT-6","kind":"llm","releaseDate":"2026-09-29","knowledgeCutoff":"2026-04-30","status":"ga","summary":"OpenAI’s mid-tier GPT-6 model for agentic coding and computer use.","description":"GPT-6.1 Sol updates GPT-6 Sol and sits below the flagship GPT-6 Astra. OpenAI targets it at agentic coding, computer use and professional work involving many documents, and describes it as token-efficient.","useCases":["coding","agents","reasoning","extraction","enterprise"],"strengths":["Agentic coding","Computer use","Token efficiency (vendor-reported)"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1050000,"maxOutputTokens":128000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":2,"output":10,"cachedInput":0.1,"note":"Prompts over 272K input tokens: $4 input / $15 output per 1M tokens"},"usage":{"tokenShare":1.37,"usageRank":16,"asOf":"2026-10-31","topTasks":[{"label":"DevOps & Config","share":4.8},{"label":"Code Generation","share":3.8},{"label":"Debugging","share":3.6}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":50.8,"codingScore":39.5,"agentScore":38.4,"mathScore":35,"toolCallingScore":23.7,"visionScore":36,"gpqa":94.4,"tauBenchAirline":70,"arenaText":1481,"arenaWebdev":1751,"eci":167.8},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":86,"links":{"website":"https://openai.com","docs":"https://platform.openai.com/docs"},"sources":["https://developers.openai.com/api/docs/pricing","https://developers.openai.com/api/docs/models","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"OpenAI","blendedPrice":4},{"id":"gemini-3.8-flash","name":"Gemini 3.8 Flash","providerId":"google","family":"Gemini 3","kind":"llm","releaseDate":"2026-09-02","knowledgeCutoff":"2026-03-31","status":"ga","summary":"Google’s most capable Flash model, with text, image, audio and video input.","description":"Gemini 3.8 Flash is the newest model in Google’s Flash line. Google reports clear improvements over 3.7 Flash in software engineering, agent tasks and multi-step reasoning. It accepts text, images, audio, video and files.","useCases":["multimodal","vision","long-context","realtime","agents","math"],"strengths":["Native audio & video input","Agentic tasks"],"technologies":["native-multimodal","reasoning"],"openWeights":false,"architecture":{},"contextWindow":1048576,"maxOutputTokens":65536,"modalities":{"input":["text","image","audio","video","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":0.75,"output":3.75,"cachedInput":0.075,"note":"Pricing through Dec 31, 2026; from Jan 1, 2027: $1.50 input / $7.50 output per 1M tokens","validUntil":"2026-12-31"},"usage":{"tokenShare":1.02,"usageRank":21,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":45.9,"codingScore":34.8,"agentScore":34.2,"toolCallingScore":23.1,"visionScore":33.7,"gpqa":93.4,"tauBenchAirline":71.3,"arenaText":1496,"arenaWebdev":1583,"eci":157.6},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":85,"links":{"website":"https://deepmind.google/technologies/gemini/","docs":"https://ai.google.dev/gemini-api/docs"},"sources":["https://ai.google.dev/gemini-api/docs/pricing","https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Google","blendedPrice":1.5},{"id":"gpt-oss-120b","name":"gpt-oss-120b","providerId":"openai","family":"gpt-oss","kind":"llm","releaseDate":"2025-08-05","status":"ga","summary":"OpenAI’s Apache-2.0 open-weight reasoning model; runs on a single 80GB GPU.","description":"gpt-oss-120b is an open-weight mixture-of-experts model from OpenAI with 117B total parameters, of which 5.1B are active per token. It is built for reasoning, agents and general production use and can be self-hosted.","useCases":["on-device","reasoning","agents","high-volume","coding"],"strengths":["Permissive Apache 2.0 license","Self-hostable","Configurable reasoning effort"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"openai/gpt-oss-120b","architecture":{"type":"MoE","totalParams":"117B","activeParams":"5.1B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.25,"usageRank":44,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":22.5,"mathScore":23,"toolCallingScore":5,"visionScore":2.4,"gpqa":72.9,"tauBenchAirline":65.4,"arenaText":1352,"eci":139.8},"benchmarksAsOf":{"reasoningScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","eci":"2026-10-31"},"popularity":82,"links":{"huggingFace":"https://huggingface.co/openai/gpt-oss-120b"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"OpenAI","blendedPrice":null},{"id":"claude-haiku-5.5","name":"Claude Haiku 5.5","providerId":"anthropic","family":"Claude 5","kind":"llm","releaseDate":"2026-10-07","knowledgeCutoff":"2026-06-01","status":"ga","summary":"Anthropic’s small, fast Claude for subagents, summaries and browser tasks.","description":"Claude Haiku 5.5 is the compact member of the Claude 5.5 line, meant for high-volume work where cost and speed matter, such as summarisation, subagents and browser automation. Anthropic says it improves coding and computer use over Claude Haiku 4.5.","useCases":["high-volume","realtime","agents","extraction","chat","coding"],"strengths":["Small and fast","Subagent work","Computer use"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1000000,"maxOutputTokens":128000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":0.1,"output":0.5,"cachedInput":0.01,"batchInput":0.05,"batchOutput":0.25,"note":"Prompts over 100K tokens: $0.50 input / $2.50 output per 1M tokens"},"usage":{"tokenShare":1.04,"usageRank":20,"asOf":"2026-10-31","topTasks":[{"label":"SQL & Database","share":6.4},{"label":"Tool Dispatch","share":5.9},{"label":"Memory Extraction","share":5.4}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":46.4,"codingScore":39.2,"agentScore":33.9,"mathScore":28.8,"toolCallingScore":20.1,"visionScore":30.1,"sweBenchPro":64.8,"gpqa":86.9,"tauBenchAirline":72.7,"arenaText":1425,"arenaWebdev":1589,"eci":160.4},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":80,"links":{"website":"https://www.anthropic.com/claude","docs":"https://docs.anthropic.com"},"sources":["https://platform.claude.com/docs/en/about-claude/pricing","https://platform.claude.com/docs/en/about-claude/models/overview","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Anthropic","blendedPrice":0.2},{"id":"text-embedding-3-small","name":"text-embedding-3-small","providerId":"openai","family":"text-embedding-3","kind":"embedding","releaseDate":"2024-01-25","status":"ga","summary":"OpenAI’s small, low-cost embedding for high-volume indexing.","description":"text-embedding-3-small is OpenAI’s low-cost embedding model and the successor to text-embedding-ada-002, producing 1536-dimension vectors.","useCases":["high-volume","research"],"strengths":["Small and inexpensive (vendor positioning)","Drop-in ada-002 replacement"],"technologies":["vector-search","rag"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":true},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":80,"links":{"website":"https://platform.openai.com/docs/guides/embeddings"},"sources":["https://openai.com/index/new-embedding-models-and-api-updates/"],"providerName":"OpenAI","blendedPrice":null},{"id":"deepseek-v4.1-flash-1017","name":"DeepSeek V4.1 Flash 1017","providerId":"deepseek","family":"DeepSeek V4","kind":"llm","releaseDate":"2026-10-17","status":"ga","summary":"October update of DeepSeek V4.1 Flash, tuned for whole-project agentic coding.","description":"DeepSeek V4.1 Flash 1017 refreshes the September V4.1 Flash weights on the same CED mixture-of-experts architecture. DeepSeek says it scores higher while finishing tasks faster with shorter outputs, after extra reinforcement learning on end-to-end projects. It is the model behind the deepseek-flash API name.","useCases":["high-volume","realtime","coding","agents","on-device"],"strengths":["Open weights (MIT)","Shorter, faster task runs","Low API price"],"technologies":["reasoning","moe"],"openWeights":true,"license":"MIT","huggingFaceId":"deepseek-ai/DeepSeek-V4.1-Flash-1017","architecture":{"type":"MoE (CED)","totalParams":"763B","activeParams":"8B in / 16B out"},"contextWindow":1000000,"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":0.3,"output":1.2,"cachedInput":0.006,"note":"Peak-hour price; off-peak (outside 01:00–04:00 and 06:00–10:00 UTC on weekdays) is half: $0.15 input / $0.60 output per 1M tokens"},"benchmarks":{"reasoningScore":48,"codingScore":40,"agentScore":39,"mathScore":31.8,"toolCallingScore":31.4,"visionScore":32.4,"arenaWebdev":1665,"eci":158.8},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":79,"links":{"website":"https://www.deepseek.com","huggingFace":"https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017"},"sources":["https://api-docs.deepseek.com/news/news261017","https://api-docs.deepseek.com/quick_start/pricing","https://llm-stats.com/leaderboards/llm-leaderboard","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"DeepSeek","blendedPrice":0.525},{"id":"gpt-6-luna","name":"GPT-6 Luna","providerId":"openai","family":"GPT-6","kind":"llm","releaseDate":"2026-09-22","knowledgeCutoff":"2026-05-18","status":"ga","summary":"OpenAI’s fast, low-cost GPT-6 tier for chat, classification and light agents.","description":"GPT-6 Luna is the lightweight tier of OpenAI’s GPT-6 family, below GPT-6 Sol. It is intended for high-volume, latency-sensitive jobs such as chat, classification and simple agent tasks.","useCases":["high-volume","chat","extraction","realtime"],"strengths":["High-volume workloads","Low latency"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1050000,"maxOutputTokens":128000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":0.1,"output":0.5,"cachedInput":0.01,"note":"Prompts over 272K input tokens: $0.20 input / $0.75 output per 1M tokens"},"usage":{"tokenShare":4.36,"usageRank":5,"asOf":"2026-10-31","topTasks":[{"label":"Customer Support","share":18.1},{"label":"DevOps & Config","share":13.1},{"label":"SQL & Database","share":12.3}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":39.3,"codingScore":29.5,"agentScore":30.5,"toolCallingScore":22.7,"visionScore":21.7,"gpqa":89.8,"tauBenchAirline":67.3,"arenaText":1441,"arenaWebdev":1579,"eci":156.8},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":78,"links":{"website":"https://openai.com","docs":"https://platform.openai.com/docs"},"sources":["https://developers.openai.com/api/docs/pricing","https://developers.openai.com/api/docs/models","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"OpenAI","blendedPrice":0.2},{"id":"text-embedding-3-large","name":"text-embedding-3-large","providerId":"openai","family":"text-embedding-3","kind":"embedding","releaseDate":"2024-01-25","status":"ga","summary":"OpenAI’s most capable embedding; the default in countless RAG stacks.","description":"text-embedding-3-large is OpenAI’s most capable embedding model for English and non-English text, with shortenable (Matryoshka-style) 3072-dimension vectors.","useCases":["research","enterprise"],"strengths":["Ubiquitous ecosystem support","Shortenable dimensions"],"technologies":["vector-search","rag"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":true},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":78,"links":{"website":"https://platform.openai.com/docs/guides/embeddings"},"sources":["https://openai.com/index/new-embedding-models-and-api-updates/"],"providerName":"OpenAI","blendedPrice":null},{"id":"nano-banana-2.1","name":"Nano Banana 2.1","providerId":"google","family":"Nano Banana","kind":"image","releaseDate":"2026-10-06","status":"ga","summary":"Google’s latest image generation and editing model.","description":"Nano Banana 2.1 is Google’s image generation and editing model, accepting text and image inputs and returning images, with strong consistency for multi-turn edits.","useCases":["image-generation","vision","multimodal"],"strengths":["Character consistency across edits","Fast generation"],"technologies":["native-multimodal"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image","text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToImage":1325,"arenaImageEdit":1427},"benchmarksAsOf":{"arenaTextToImage":"2026-10-26","arenaImageEdit":"2026-10-27"},"popularity":76,"links":{},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Google","blendedPrice":null},{"id":"gpt-5.4-image-2","name":"GPT-5.4 Image 2","providerId":"openai","family":"GPT Image","kind":"image","releaseDate":"2026-04-21","status":"ga","summary":"OpenAI’s native image generation and editing model.","description":"GPT-5.4 Image 2 generates and edits images from text, image and file inputs, with strong text rendering and instruction following.","useCases":["image-generation","vision","multimodal"],"strengths":["Accurate text in images","Conversational editing"],"technologies":["native-multimodal"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image","file"],"output":["image","text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":74,"links":{},"sources":[],"providerName":"OpenAI","blendedPrice":null},{"id":"gpt-oss-20b","name":"gpt-oss-20b","providerId":"openai","family":"gpt-oss","kind":"llm","releaseDate":"2025-08-05","status":"ga","summary":"Compact open-weight reasoning model that runs on consumer hardware.","description":"gpt-oss-20b is OpenAI’s smaller open-weight model: a 21B-parameter mixture of experts with 3.6B active parameters, released under Apache 2.0 for low-latency and local use.","useCases":["on-device","high-volume","realtime","extraction"],"strengths":["Runs on a laptop GPU","Apache 2.0","Self-hostable"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"openai/gpt-oss-20b","architecture":{"type":"MoE","totalParams":"21B","activeParams":"3.6B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"asOf":"2026-10-31","topTasks":[{"label":"Memory Extraction","share":4.5}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":14.9,"mathScore":16.6,"toolCallingScore":-2.7,"visionScore":-2,"gpqa":64.7,"tauBenchAirline":53.6,"arenaText":1317,"eci":137.8},"benchmarksAsOf":{"reasoningScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","eci":"2026-10-31"},"popularity":72,"links":{"huggingFace":"https://huggingface.co/openai/gpt-oss-20b"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"OpenAI","blendedPrice":null},{"id":"kimi-k3","name":"Kimi K3","providerId":"moonshot","family":"Kimi K","kind":"llm","releaseDate":"2026-07-16","status":"ga","summary":"Moonshot AI’s 2.8T-parameter open multimodal reasoning model.","description":"Kimi K3 is an open-weight multimodal reasoning model from Moonshot AI with 2.8T parameters, intended for complex coding, knowledge work and long agent workflows. It takes text, images and video.","useCases":["reasoning","coding","agents","math","research"],"strengths":["Frontier-scale open weights","Video input"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Modified MIT","huggingFaceId":"moonshotai/Kimi-K3","architecture":{"type":"MoE","totalParams":"2.8T"},"contextWindow":1048576,"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":3,"output":15,"cachedInput":0.3},"usage":{"tokenShare":1.12,"usageRank":19,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":50.4,"codingScore":40.7,"agentScore":37.6,"mathScore":41.5,"toolCallingScore":31.3,"visionScore":35.9,"gpqa":91.8,"tauBenchAirline":73.3,"arenaText":1488,"arenaWebdev":1656,"eci":158.3},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":71,"links":{"website":"https://www.moonshot.ai","huggingFace":"https://huggingface.co/moonshotai/Kimi-K3"},"sources":["https://platform.kimi.ai/docs/pricing/chat","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Moonshot AI","blendedPrice":6},{"id":"claude-fable-5.1","name":"Claude Fable 5.1","providerId":"anthropic","family":"Claude 5","kind":"llm","releaseDate":"2026-09-01","knowledgeCutoff":"2026-06-01","status":"ga","summary":"Anthropic’s premium model for long agent workflows, refactors and visual front-end work.","description":"Claude Fable 5.1 is an update to Claude Fable 5. Anthropic highlights gains in agentic coding and long-running workflows, including lengthy refactors, front-end and visual design, and document-heavy knowledge work.","useCases":["agents","coding","research","writing","enterprise"],"strengths":["Long-running autonomous work","Front-end & visual design","Knowledge work"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1000000,"maxOutputTokens":128000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":10,"output":50,"cachedInput":0.25,"batchInput":5,"batchOutput":25},"usage":{"tokenShare":0.05,"usageRank":58,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":50.4,"codingScore":35.5,"agentScore":38.7,"mathScore":44.4,"toolCallingScore":28.1,"visionScore":41.8,"gpqa":89.9,"tauBenchAirline":79.2,"arenaText":1504,"arenaWebdev":1746,"eci":166.4},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":70,"links":{"website":"https://www.anthropic.com/claude","docs":"https://docs.anthropic.com"},"sources":["https://platform.claude.com/docs/en/about-claude/pricing","https://platform.claude.com/docs/en/about-claude/models/overview","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Anthropic","blendedPrice":20},{"id":"gpt-6.1-terra","name":"GPT-6.1 Terra","providerId":"openai","family":"GPT-6","kind":"llm","releaseDate":"2026-10-26","knowledgeCutoff":"2026-06-01","status":"ga","summary":"OpenAI’s fast, lower-cost GPT-6 model for computer use, visual reasoning and writing.","description":"GPT-6.1 Terra is the efficient tier of OpenAI’s GPT-6 line, below GPT-6.1 Sol and GPT-6 Astra. OpenAI aims it at visual reasoning, precise computer use and everyday writing, at half the per-token price of GPT-6.1 Sol. Reasoning effort runs from none to max.","useCases":["agents","vision","writing","chat"],"strengths":["Half GPT-6.1 Sol’s price","Computer use","Image input"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1050000,"maxOutputTokens":128000,"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":1,"output":5,"cachedInput":0.05,"note":"Prompts over 272K input tokens: $2 input / $7.50 output per 1M tokens"},"benchmarks":{"reasoningScore":49,"codingScore":37.1,"agentScore":38.8,"toolCallingScore":33.3,"arenaWebdev":1673,"eci":168.1},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","toolCallingScore":"2026-10-31","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":70,"links":{"website":"https://openai.com"},"sources":["https://developers.openai.com/api/docs/pricing","https://developers.openai.com/api/docs/models/gpt-6.1-terra","https://developers.openai.com/api/docs/changelog","https://llm-stats.com/leaderboards/llm-leaderboard","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"OpenAI","blendedPrice":2},{"id":"llama-4-maverick","name":"Llama 4 Maverick","providerId":"meta","family":"Llama 4","kind":"llm","releaseDate":"2025-04-05","status":"ga","summary":"Meta’s 128-expert open multimodal MoE model.","description":"Llama 4 Maverick is Meta’s larger Llama 4 model: a mixture of experts with 128 experts, 400B total and 17B active parameters, accepting text and images.","useCases":["on-device","chat","vision","multimodal"],"strengths":["Huge open ecosystem","Fine-tuning friendly"],"technologies":["moe"],"openWeights":true,"license":"Llama 4 Community License","huggingFaceId":"meta-llama/Llama-4-Maverick-17B-128E-Instruct","architecture":{"type":"MoE","totalParams":"400B","activeParams":"17B"},"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":false,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":13.8,"codingScore":2.9,"mathScore":20.3,"visionScore":12.3,"gpqa":66.6,"tauBenchAirline":42.7,"arenaText":1327,"eci":132.3},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","mathScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","eci":"2026-10-31"},"popularity":70,"links":{"huggingFace":"https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Meta","blendedPrice":null},{"id":"gpt-image-2.5","name":"GPT Image 2.5","providerId":"openai","family":"GPT Image","kind":"image","releaseDate":"2026-09-08","status":"ga","summary":"OpenAI’s newest image model, in Sunburst and Flare variants.","description":"GPT Image 2.5 is OpenAI’s latest image generation and editing model, available in Sunburst and Flare variants.","useCases":["image-generation","writing"],"strengths":["Two quality variants","Strong text rendering"],"technologies":["native-multimodal"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToImage":1427,"arenaImageEdit":1517},"benchmarksAsOf":{"arenaTextToImage":"2026-10-26","arenaImageEdit":"2026-10-27"},"popularity":70,"links":{"website":"https://openai.com"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"OpenAI","blendedPrice":null},{"id":"deepseek-v4-pro","name":"DeepSeek V4 Pro","providerId":"deepseek","family":"DeepSeek V4","kind":"llm","releaseDate":"2026-08-12","status":"ga","summary":"DeepSeek’s large open MoE, GA release (0813).","description":"DeepSeek V4 Pro 0813 is the generally available release of DeepSeek’s 1.6T-parameter mixture-of-experts model, published with MIT-licensed weights.","useCases":["chat","coding","high-volume","on-device"],"strengths":["Open weights (MIT)","Frontier-scale MoE"],"technologies":["reasoning","moe"],"openWeights":true,"license":"MIT","huggingFaceId":"deepseek-ai/DeepSeek-V4-Pro-0813","architecture":{"type":"MoE","totalParams":"1.6T"},"contextWindow":1000000,"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":1.32,"output":3.96,"cachedInput":0.044,"note":"Peak-hour price; off-peak (outside 01:00–04:00 and 06:00–10:00 UTC on weekdays) is half: $0.66 input / $1.98 output per 1M tokens"},"usage":{"tokenShare":0.92,"usageRank":22,"asOf":"2026-10-31","topTasks":[{"label":"Roleplay & Fiction","share":3.2}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":48,"codingScore":38.7,"agentScore":35.3,"mathScore":39.1,"toolCallingScore":28.4,"visionScore":33.4,"gpqa":86.9,"tauBenchAirline":76.8,"arenaText":1465,"arenaWebdev":1582,"eci":155.9},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":69,"links":{"huggingFace":"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://api-docs.deepseek.com/quick_start/pricing","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"DeepSeek","blendedPrice":1.98},{"id":"gemma-4-31b","name":"Gemma 4 31B","providerId":"google","family":"Gemma 4","kind":"llm","releaseDate":"2026-04-02","knowledgeCutoff":"2025-01-01","status":"ga","summary":"Open-weight dense multimodal model from Google DeepMind.","description":"Gemma 4 31B is Google DeepMind’s largest dense Gemma 4 model (30.7B parameters). It reads text, images and video, can switch thinking on or off and supports function calling.","useCases":["on-device","vision","chat","extraction"],"strengths":["Single-GPU deployable","Multimodal open model","Function calling"],"technologies":["reasoning","transformer"],"openWeights":true,"license":"Gemma Terms of Use","huggingFaceId":"google/gemma-4-31B-it","architecture":{"type":"Dense","totalParams":"30.7B"},"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.27,"usageRank":41,"asOf":"2026-10-31","topTasks":[{"label":"Roleplay & Fiction","share":3}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":33,"agentScore":13.3,"mathScore":29.2,"toolCallingScore":13.3,"visionScore":19.6,"gpqa":81.7,"tauBenchAirline":75.3,"arenaText":1452,"arenaWebdev":1369,"eci":143.1},"benchmarksAsOf":{"reasoningScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":68,"links":{"huggingFace":"https://huggingface.co/google/gemma-4-31B-it"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Google","blendedPrice":null},{"id":"qwen3.8-flash","name":"Qwen3.8 Flash","providerId":"alibaba","family":"Qwen3.8","kind":"llm","releaseDate":"2026-08-26","status":"ga","summary":"Open multimodal reasoning model for coding, desktop interaction and long video.","description":"Qwen3.8 Flash is an Apache-licensed multimodal reasoning model from the Qwen team, aimed at coding help, agent workflows, document and codebase analysis, operating desktop apps and understanding long videos.","useCases":["coding","vision","agents","high-volume","multimodal"],"strengths":["Apache 2.0 weights","Long-video analysis"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"Qwen/Qwen3.8-Flash-Next","architecture":{"type":"MoE","totalParams":"125B"},"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":0.15,"output":0.47},"usage":{"tokenShare":0.33,"usageRank":37,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":48,"codingScore":33.6,"agentScore":33.1,"mathScore":32.6,"toolCallingScore":30.1,"visionScore":34.1,"sweBenchPro":62.5,"gpqa":88.6,"tauBenchAirline":71.3},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30"},"popularity":67,"links":{"huggingFace":"https://huggingface.co/Qwen/Qwen3.8-Flash-Next"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://www.alibabacloud.com/help/en/model-studio/model-pricing","https://openrouter.ai/rankings"],"providerName":"Alibaba (Qwen)","blendedPrice":0.23},{"id":"gemini-3.5-flash-lite","name":"Gemini 3.5 Flash Lite","providerId":"google","family":"Gemini 3","kind":"llm","releaseDate":"2026-07-21","knowledgeCutoff":"2026-03-31","status":"ga","summary":"Google’s efficient multimodal model for focused subagent tasks.","description":"Gemini 3.5 Flash Lite is a lightweight Gemini model with improved agent abilities, meant for subagents that each handle a narrow task inside a larger multi-agent system.","useCases":["high-volume","multimodal","extraction","agents"],"strengths":["Full multimodal input","Subagent workloads"],"technologies":["native-multimodal","reasoning"],"openWeights":false,"architecture":{},"contextWindow":1048576,"maxOutputTokens":65536,"modalities":{"input":["text","image","audio","video","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":0.3,"output":2.5,"cachedInput":0.03},"usage":{"asOf":"2026-10-31","topTasks":[{"label":"Classification","share":2.5},{"label":"Q&A & Knowledge","share":2.2}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":25.4,"codingScore":15.6,"agentScore":15.8,"toolCallingScore":3.5,"visionScore":19.2,"sweBenchPro":54.2,"gpqa":84.4,"tauBenchAirline":78.3,"arenaText":1456,"arenaWebdev":1434,"eci":145},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":66,"links":{"docs":"https://ai.google.dev/gemini-api/docs"},"sources":["https://ai.google.dev/gemini-api/docs/pricing","https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Google","blendedPrice":0.85},{"id":"deepseek-v4.1-flash","name":"DeepSeek V4.1 Flash","providerId":"deepseek","family":"DeepSeek V4","kind":"llm","releaseDate":"2026-09-10","status":"ga","summary":"Open-weight MoE built on DeepSeek’s new encoder-decoder architecture.","description":"DeepSeek V4.1 Flash is a sparse mixture-of-experts model and the first on DeepSeek’s Causal Encoder-Decoder (CED) design, which activates 8B parameters when reading input and 16B when generating. This is the original September checkpoint; DeepSeek’s API now serves the 1017 update instead.","useCases":["high-volume","realtime","coding","on-device","agents"],"strengths":["New CED architecture","Open weights (MIT)"],"technologies":["reasoning","moe"],"openWeights":true,"license":"MIT","huggingFaceId":"deepseek-ai/DeepSeek-V4.1-Flash","architecture":{"type":"MoE (CED)","totalParams":"763B","activeParams":"8B in / 16B out"},"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":24.87,"usageRank":1,"asOf":"2026-10-31","topTasks":[{"label":"Security Audit","share":49.7},{"label":"Code Generation","share":32.4},{"label":"DevOps & Config","share":31.7}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":47.2,"codingScore":40.8,"agentScore":38.4,"mathScore":35.3,"toolCallingScore":32.3,"visionScore":30.7,"gpqa":90.7,"tauBenchAirline":75.8,"arenaText":1476,"arenaWebdev":1619,"eci":156.1},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":66,"links":{"website":"https://www.deepseek.com","huggingFace":"https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"DeepSeek","blendedPrice":null},{"id":"glm-5.3","name":"GLM-5.3","providerId":"zai","family":"GLM-5","kind":"llm","releaseDate":"2026-08-18","status":"ga","summary":"Open reasoning model for complex software engineering and long-horizon agents.","description":"GLM-5.3 is Z.ai’s large MIT-licensed reasoning model, a 753B-parameter mixture of experts designed for hard software engineering and long agent tasks.","useCases":["coding","agents","reasoning","long-context"],"strengths":["Open weights (MIT)","Long agent tasks"],"technologies":["reasoning","moe"],"openWeights":true,"license":"MIT","huggingFaceId":"zai-org/GLM-5.3","architecture":{"type":"MoE","totalParams":"753B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":1.4,"output":4.4,"cachedInput":0.26},"usage":{"tokenShare":2.51,"usageRank":11,"asOf":"2026-10-31","topTasks":[{"label":"Repo Scanning","share":3.9},{"label":"Security Audit","share":3.7},{"label":"Debugging","share":3}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":50.1,"codingScore":39.9,"agentScore":37.2,"mathScore":41.9,"toolCallingScore":30.9,"visionScore":34.8,"gpqa":85.2,"tauBenchAirline":76.8,"arenaText":1481,"arenaWebdev":1623,"eci":156.2},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":66,"links":{"website":"https://z.ai","huggingFace":"https://huggingface.co/zai-org/GLM-5.3"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://docs.z.ai/guides/overview/pricing","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Z.ai","blendedPrice":2.15},{"id":"qwen3.6-35b-a3b","name":"Qwen3.6 35B A3B","providerId":"alibaba","family":"Qwen3.6","kind":"llm","releaseDate":"2026-04-26","status":"ga","summary":"Open MoE with just 3B active parameters — popular for local inference.","description":"Qwen3.6 35B A3B is an open-weight mixture of experts that activates only 3B of 35B parameters, which keeps it quick on consumer GPUs and Apple silicon. It also accepts images and video.","useCases":["on-device","high-volume","realtime","vision"],"strengths":["Runs fast locally","Multimodal"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"Qwen/Qwen3.6-35B-A3B","architecture":{"type":"MoE","totalParams":"35B","activeParams":"3B"},"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.12,"usageRank":55,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":32.4,"codingScore":19.2,"agentScore":11.8,"mathScore":30.6,"toolCallingScore":11.2,"visionScore":23.2,"sweBenchPro":49.5,"gpqa":79.4,"tauBenchAirline":72.3,"eci":143.8},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","eci":"2026-10-31"},"popularity":65,"links":{"huggingFace":"https://huggingface.co/Qwen/Qwen3.6-35B-A3B"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://epoch.ai/benchmarks"],"providerName":"Alibaba (Qwen)","blendedPrice":null},{"id":"grok-4.7","name":"Grok 4.7","providerId":"xai","family":"Grok 4","kind":"llm","releaseDate":"2026-09-21","knowledgeCutoff":"2026-06-01","status":"ga","summary":"SpaceXAI’s flagship for long software-engineering tasks and checking its own work.","description":"Grok 4.7 follows Grok 4.6 as SpaceXAI’s flagship for coding, agents and knowledge work. SpaceXAI emphasises long software-engineering runs and the model’s habit of verifying its own output.","useCases":["coding","agents","reasoning","research"],"strengths":["Self-verification","Long software-engineering runs"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":500000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":2,"output":6,"cachedInput":0.5,"note":"Prompts of 200K+ tokens: $4 input / $12 output per 1M tokens"},"usage":{"tokenShare":0.18,"usageRank":49,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":41.9,"codingScore":35.6,"agentScore":33.8,"toolCallingScore":19.9,"arenaText":1441,"arenaWebdev":1637,"eci":154.6},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","toolCallingScore":"2026-10-31","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":64,"links":{"website":"https://x.ai","docs":"https://docs.x.ai"},"sources":["https://docs.x.ai/docs/models","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"xAI (SpaceXAI)","blendedPrice":3},{"id":"eleven-v4","name":"Eleven v4","providerId":"elevenlabs","family":"Eleven","kind":"speech","releaseDate":"2026-09-28","status":"ga","summary":"ElevenLabs’ newest text-to-speech model, with a Turbo variant.","description":"Eleven v4 is the latest ElevenLabs text-to-speech model, offered alongside a faster, cheaper Turbo variant.","useCases":["multimodal","realtime","writing"],"strengths":["Turbo variant","Large voice library"],"technologies":["realtime-voice"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["audio"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":64,"links":{"website":"https://elevenlabs.io"},"sources":[],"providerName":"ElevenLabs","blendedPrice":null},{"id":"qwen3.8-max","name":"Qwen3.8 Max","providerId":"alibaba","family":"Qwen3.8","kind":"llm","releaseDate":"2026-09-03","status":"ga","summary":"Alibaba’s 2.4T-parameter proprietary flagship with video understanding.","description":"Qwen3.8 Max is Alibaba’s top proprietary Qwen model, a 2.4-trillion-parameter mixture of experts that reads text, images and video and writes text.","useCases":["reasoning","multimodal","coding","research"],"strengths":["Video input","Strong multilingual (esp. Chinese)"],"technologies":["reasoning","moe"],"openWeights":false,"architecture":{"type":"MoE","totalParams":"2.4T","activeParams":"95B"},"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":2,"output":6},"usage":{"tokenShare":0.02,"usageRank":62,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":49.3,"codingScore":37.5,"agentScore":35.9,"mathScore":33.7,"toolCallingScore":29.7,"visionScore":38,"sweBenchPro":67.7,"arenaWebdev":1676,"eci":156.2},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":63,"links":{"website":"https://qwen.ai"},"sources":["https://www.alibabacloud.com/help/en/model-studio/model-pricing","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Alibaba (Qwen)","blendedPrice":3},{"id":"gpt-5.6-terra","name":"GPT-5.6 Terra","providerId":"openai","family":"GPT-5.6","kind":"llm","releaseDate":"2026-07-09","knowledgeCutoff":"2026-02-16","status":"ga","summary":"Balanced GPT-5.6 model for everyday coding and reasoning.","description":"GPT-5.6 Terra is the middle tier of OpenAI’s GPT-5.6 series, between the Sol and Luna tiers, aimed at routine coding, reasoning and agent work.","useCases":["coding","chat","reasoning","agents"],"strengths":["Middle tier of GPT-5.6","Mature ecosystem support"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1050000,"maxOutputTokens":128000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":2,"output":12,"cachedInput":0.2,"note":"Prompts over 272K input tokens: $4 input / $18 output per 1M tokens"},"usage":{"tokenShare":0.02,"usageRank":66,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":47,"codingScore":39,"agentScore":35.9,"mathScore":31.7,"toolCallingScore":27.1,"visionScore":27.7,"sweBenchPro":63.4,"gpqa":88.4,"tauBenchAirline":72.3,"arenaText":1467,"arenaWebdev":1520,"eci":160.1},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":62,"links":{"website":"https://openai.com"},"sources":["https://developers.openai.com/api/docs/pricing","https://developers.openai.com/api/docs/models/gpt-5.6-terra","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"OpenAI","blendedPrice":4.5},{"id":"muse-spark-open","name":"Muse Spark Open","providerId":"meta","family":"Muse","kind":"llm","releaseDate":"2026-10-15","status":"ga","summary":"Meta’s open-weight 816B MoE, tuned for agentic coding.","description":"Muse Spark Open is an open-weight Muse model from Meta: a mixture of experts with 816B total and 42B active parameters, trained mainly for agentic coding, tool use and understanding large repositories. It accepts text, images, video, audio and files.","useCases":["coding","agents","reasoning","long-context"],"strengths":["Open weights at 816B scale","Agentic coding focus"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Muse Model License","huggingFaceId":"meta-models/Muse-Spark-Open-816B-A42B","architecture":{"type":"MoE","totalParams":"816B","activeParams":"42B"},"modalities":{"input":["text","image","video","audio","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"reasoningScore":45.1,"codingScore":34,"agentScore":31.1,"mathScore":33,"toolCallingScore":24.8,"visionScore":31.3,"arenaWebdev":1526,"eci":155.1},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":62,"links":{"huggingFace":"https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Meta","blendedPrice":null},{"id":"muse-spark-1.3","name":"Muse Spark 1.3","providerId":"meta","family":"Muse","kind":"llm","releaseDate":"2026-09-02","status":"ga","summary":"Meta’s multimodal reasoning model for long-running agents and coding.","description":"Muse Spark 1.3 is a Meta reasoning model for long agent, multi-agent and coding workflows, designed to keep track of details over extended tasks. It reads text, images, files and video.","useCases":["agents","coding","reasoning","multimodal","long-context"],"strengths":["Long-running agents","Video input"],"technologies":["native-multimodal","reasoning"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image","file","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":1.51,"usageRank":15,"asOf":"2026-10-31","topTasks":[{"label":"Frontend & UI","share":6.1},{"label":"Code Generation","share":2.4}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":50.3,"codingScore":38.6,"agentScore":37.9,"toolCallingScore":31.1,"visionScore":27,"arenaText":1495,"arenaWebdev":1656,"eci":157.4},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":62,"links":{"website":"https://ai.meta.com"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Meta","blendedPrice":null},{"id":"claude-opus-5","name":"Claude Opus 5","providerId":"anthropic","family":"Claude 5","kind":"llm","releaseDate":"2026-07-24","knowledgeCutoff":"2026-05-01","status":"ga","summary":"Previous-generation Opus; still a strong reasoning and coding model.","description":"Claude Opus 5 was Anthropic’s flagship before Opus 5.5. It remains available for teams pinned to it, with a 1M-token context window.","useCases":["coding","agents","reasoning","research"],"strengths":["Proven in production","Deep reasoning"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1000000,"maxOutputTokens":128000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":5,"output":25,"cachedInput":0.5,"batchInput":2.5,"batchOutput":12.5},"usage":{"tokenShare":0.36,"usageRank":34,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":53.5,"codingScore":39.5,"agentScore":38.7,"mathScore":43.4,"toolCallingScore":25.2,"visionScore":38.4,"gpqa":87.8,"tauBenchAirline":78.9,"arenaText":1492,"arenaWebdev":1691,"eci":164.5},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":60,"links":{"website":"https://www.anthropic.com/claude"},"sources":["https://platform.claude.com/docs/en/about-claude/pricing","https://platform.claude.com/docs/en/models/opus-5/overview","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Anthropic","blendedPrice":10},{"id":"kimi-k3.1-pro","name":"Kimi K3.1 Pro","providerId":"moonshot","family":"Kimi K","kind":"llm","releaseDate":"2026-10-23","status":"ga","summary":"Post-trained Kimi K3 for coding and agent work, with a 1M-token context.","description":"Kimi K3.1 Pro is Moonshot AI’s post-trained successor to Kimi K3, aimed at coding, financial research and other long agent tasks. It reads text, images and video, always reasons before answering, and costs less per token than K3. It is on the Kimi API and apps now; Moonshot says the full weights will be published by November 4, 2026.","useCases":["coding","agents","reasoning","research","long-context"],"strengths":["1M-token context","Video input","Cheaper than Kimi K3"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1048576,"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":2,"output":10,"cachedInput":0.2,"note":"K3.1 Pro Highspeed (2–3× faster): $4 input / $20 output per 1M tokens"},"usage":{"tokenShare":0.28,"usageRank":40,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":54.3,"codingScore":42.7,"agentScore":41.6,"mathScore":38.6,"toolCallingScore":30.4,"visionScore":40.5,"gpqa":90.4,"arenaWebdev":1695},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","arenaWebdev":"2026-10-28"},"popularity":60,"links":{"website":"https://www.moonshot.ai"},"sources":["https://www.kimi.ai/resources/kimi-k3-1-pro-pricing","https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart","https://forum.moonshot.ai/t/kimi-k3-1-pro-has-launched-start-exploring-now/629","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Moonshot AI","blendedPrice":4},{"id":"veo-3.1","name":"Veo 3.1","providerId":"google","family":"Veo","kind":"video","releaseDate":"2026-01-30","status":"ga","summary":"Google’s Veo video model with native audio; powers Flow.","description":"Veo 3.1 generates video with synchronised audio from text and images and powers Google’s Flow filmmaking tool. Fast and Lite variants trade some quality for speed and cost.","useCases":["multimodal"],"strengths":["Native audio","Fast & Lite tiers"],"technologies":["diffusion"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["video","audio"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToVideo":1364,"arenaImageToVideo":1398},"benchmarksAsOf":{"arenaTextToVideo":"2026-09-22","arenaImageToVideo":"2026-09-22"},"popularity":60,"links":{"website":"https://deepmind.google/models/veo/"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Google","blendedPrice":null},{"id":"llama-4-scout","name":"Llama 4 Scout","providerId":"meta","family":"Llama 4","kind":"llm","releaseDate":"2025-04-05","status":"ga","summary":"Efficient open Llama 4 model with a very long native context.","description":"Llama 4 Scout is the smaller Llama 4 model, activating 17B of its 109B parameters across 16 experts. Meta designed it for multimodal input and very long documents.","useCases":["long-context","on-device","high-volume"],"strengths":["Very long native context (vendor-reported)","Cheap to self-host"],"technologies":["moe"],"openWeights":true,"license":"Llama 4 Community License","huggingFaceId":"meta-llama/Llama-4-Scout-17B-16E-Instruct","architecture":{"type":"MoE","totalParams":"109B","activeParams":"17B"},"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":false,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"reasoningScore":6.1,"codingScore":0.3,"mathScore":12.8,"visionScore":9.7,"arenaText":1321,"eci":129.6},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","mathScore":"2026-10-31","visionScore":"2026-10-31","arenaText":"2026-10-28","eci":"2026-10-31"},"popularity":58,"links":{"huggingFace":"https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Meta","blendedPrice":null},{"id":"glm-5.3-flash","name":"GLM-5.3 Flash","providerId":"zai","family":"GLM-5","kind":"llm","releaseDate":"2026-08-26","status":"ga","summary":"Native multimodal open model with hybrid sparse/linear attention for long inputs.","description":"GLM-5.3 Flash is Z.ai’s efficient multimodal model for coding and long agent tasks. It combines sparse and linear attention so that quality holds up on long inputs.","useCases":["coding","agents","high-volume","long-context","vision"],"strengths":["Hybrid sparse/linear attention","Open weights (MIT)"],"technologies":["efficient-attention","native-multimodal","reasoning","moe"],"openWeights":true,"license":"MIT","huggingFaceId":"zai-org/GLM-5.3-Flash","architecture":{"type":"MoE","totalParams":"320B"},"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":0.15,"output":0.5,"cachedInput":0.03},"usage":{"tokenShare":7.21,"usageRank":2,"asOf":"2026-10-31","topTasks":[{"label":"Data Transformation","share":45.9},{"label":"Multi-step Planning","share":36.6},{"label":"Workflow Execution","share":25.6}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":47.4,"codingScore":32.4,"agentScore":35,"mathScore":37.3,"toolCallingScore":29.6,"visionScore":31.3,"gpqa":84.4,"tauBenchAirline":75.6,"arenaText":1477,"arenaWebdev":1607,"eci":153.2},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":58,"links":{"huggingFace":"https://huggingface.co/zai-org/GLM-5.3-Flash"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://docs.z.ai/guides/overview/pricing","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Z.ai","blendedPrice":0.238},{"id":"muse-wonder","name":"Muse Wonder","providerId":"meta","family":"Muse","kind":"llm","releaseDate":"2026-10-12","status":"ga","summary":"Meta’s first frontier-class model, built for sustained agentic work.","description":"Muse Wonder is Meta Superintelligence Labs’ most capable API model, released October 12, 2026 for long, complex personal and professional agent tasks. It takes text, images, audio, video and PDFs. Meta highlights Active Memory, which carries information across conversations, and Active Focus, which recovers from errors in multi-step tasks. It is available through the Meta Model API and Muse Code.","useCases":["agents","coding","reasoning","multimodal","long-context"],"strengths":["Active Memory and Active Focus","Audio and video input"],"technologies":["native-multimodal","reasoning"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image","file","audio","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.01,"usageRank":70,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"arenaText":1507,"arenaWebdev":1652,"eci":161.7},"benchmarksAsOf":{"arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":58,"links":{"website":"https://ai.meta.com"},"sources":["https://research.meta.ai/blog/muse-wonder","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Meta","blendedPrice":null},{"id":"qwen3-embedding-8b","name":"Qwen3-Embedding-8B","providerId":"alibaba","family":"Qwen3 Embedding","kind":"embedding","releaseDate":"2025-06-05","status":"ga","summary":"Leading open-weight multilingual embedding model.","description":"Qwen3-Embedding-8B is part of the Qwen3 Embedding series for text embedding and ranking, inheriting Qwen3’s multilingual and long-text understanding. Apache 2.0 weights can be self-hosted.","useCases":["on-device","research","enterprise"],"strengths":["Open weights (Apache 2.0)","32K-token inputs","Multilingual"],"technologies":["vector-search","rag","transformer"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"Qwen/Qwen3-Embedding-8B","architecture":{"type":"Dense","totalParams":"8B"},"modalities":{"input":["text"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":58,"links":{"website":"https://qwenlm.github.io/blog/qwen3-embedding/","huggingFace":"https://huggingface.co/Qwen/Qwen3-Embedding-8B"},"sources":[],"providerName":"Alibaba (Qwen)","blendedPrice":null},{"id":"flux-3-image","name":"FLUX 3 Image","providerId":"black-forest-labs","family":"FLUX","kind":"image","releaseDate":"2026-10-01","status":"ga","summary":"Black Forest Labs’ newest FLUX image model.","description":"FLUX 3 Image is the latest text-to-image model in Black Forest Labs’ FLUX family.","useCases":["image-generation"],"strengths":["FLUX family","Widely available via API partners"],"technologies":["diffusion"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":58,"links":{"website":"https://bfl.ai"},"sources":[],"providerName":"Black Forest Labs","blendedPrice":null},{"id":"mimo-v2.6-pro","name":"MiMo-V2.6-Pro","providerId":"xiaomi","family":"MiMo","kind":"llm","releaseDate":"2026-09-21","status":"ga","summary":"Xiaomi’s 1T-parameter open flagship with omni-modal input.","description":"MiMo-V2.6-Pro is Xiaomi’s flagship foundation model, at just over 1T parameters. It reads text, images, audio and video, and its weights are MIT-licensed.","useCases":["reasoning","coding","multimodal","high-volume","agents"],"strengths":["Omni-modal input","Open weights (MIT)"],"technologies":["reasoning","moe"],"openWeights":true,"license":"MIT","huggingFaceId":"XiaomiMiMo/MiMo-V2.6-Pro-RL","architecture":{"type":"MoE","totalParams":"1.02T"},"modalities":{"input":["text","image","audio","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.81,"usageRank":23,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":45.3,"codingScore":39,"agentScore":36.4,"toolCallingScore":29.7,"visionScore":27.7,"gpqa":88.7,"tauBenchAirline":76.8,"arenaText":1480,"arenaWebdev":1625},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28"},"popularity":57,"links":{"huggingFace":"https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Xiaomi","blendedPrice":null},{"id":"qwen3.8-27b","name":"Qwen3.8 27B","providerId":"alibaba","family":"Qwen3.8","kind":"llm","releaseDate":"2026-08-14","status":"ga","summary":"Dense 27B open vision-language model with adjustable thinking.","description":"Qwen3.8 27B is a dense, Apache-licensed vision-language model from the Qwen team for coding, research, multimodal tasks and long agent runs. Its thinking can be turned off or scaled up to an extra-high effort setting.","useCases":["coding","agents","vision","on-device"],"strengths":["Apache 2.0 weights","Adjustable thinking effort"],"technologies":["reasoning","transformer"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"Qwen/Qwen3.8-27B","architecture":{"type":"Dense","totalParams":"27B"},"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.27,"usageRank":42,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":43.5,"codingScore":29.9,"agentScore":29.7,"mathScore":30.8,"toolCallingScore":21.3,"visionScore":33.1,"sweBenchPro":61.7,"gpqa":81.8,"tauBenchAirline":78,"arenaText":1436,"arenaWebdev":1590,"eci":151.8},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":56,"links":{"huggingFace":"https://huggingface.co/Qwen/Qwen3.8-27B"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Alibaba (Qwen)","blendedPrice":null},{"id":"mistral-small-4","name":"Mistral Small 4","providerId":"mistral","family":"Mistral Small","kind":"llm","releaseDate":"2026-03-16","status":"ga","summary":"Apache-licensed open model unifying reasoning, vision and coding.","description":"Mistral Small 4 folds several earlier Mistral models into one 119B-parameter system that handles reasoning, images and code, released under Apache 2.0.","useCases":["on-device","high-volume","vision","extraction","realtime"],"strengths":["Apache 2.0","Reasoning, vision and coding in one"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"mistralai/Mistral-Small-4-119B-2603","architecture":{"type":"MoE","totalParams":"119B","activeParams":"6.5B"},"contextWindow":256000,"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":0.15,"output":0.6},"benchmarks":{},"benchmarksAsOf":{},"popularity":56,"links":{"huggingFace":"https://huggingface.co/mistralai/Mistral-Small-4-119B-2603"},"sources":["https://docs.mistral.ai/models/mistral-small-4-0-26-03"],"providerName":"Mistral AI","blendedPrice":0.262},{"id":"gemini-4-argon","name":"Gemini 4 Argon","providerId":"google","family":"Gemini 4","kind":"llm","releaseDate":"2026-09-30","status":"ga","summary":"Google’s flagship reasoning model for software engineering, science and long-running agents.","description":"Google announced Gemini 4 Argon on September 30, 2026, first for vetted cyber-defense organisations, and opened it to developers and enterprises on October 13, 2026 through the Gemini API, AI Studio and the Gemini app. Google lists a shared 1M-token budget for input and output, and introductory launch pricing until November 30, 2026.","useCases":["reasoning","coding","research","agents"],"strengths":["Google’s flagship","Text, image, audio and video input"],"technologies":["native-multimodal","reasoning"],"openWeights":false,"architecture":{},"contextWindow":1048576,"maxOutputTokens":65536,"modalities":{"input":["text","image","audio","video","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":2,"output":10,"cachedInput":0.1,"note":"Introductory pricing through Nov 30, 2026; from Dec 1, 2026: $4 input / $20 output per 1M tokens","validUntil":"2026-11-30"},"usage":{"tokenShare":0.33,"usageRank":36,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":52,"codingScore":43.2,"agentScore":41.5,"toolCallingScore":32.4,"visionScore":36.2,"gpqa":96,"tauBenchAirline":78.7,"arenaText":1520,"arenaWebdev":1679,"eci":166.1},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":55,"links":{"website":"https://deepmind.google/technologies/gemini/"},"sources":["https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/","https://ai.google.dev/gemini-api/docs/pricing","https://ai.google.dev/gemini-api/docs/models/gemini-4-argon","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Google","blendedPrice":4},{"id":"grok-4.20","name":"Grok 4.20","providerId":"xai","family":"Grok 4","kind":"llm","releaseDate":"2026-03-31","status":"ga","summary":"Fast reasoning model with a 1M-token context window and strict prompt adherence.","description":"Grok 4.20 is a reasoning model with fast agentic tool calling, a low hallucination rate, strict prompt adherence and a 1M-token context window.","useCases":["long-context","agents","realtime","extraction"],"strengths":["1M-token context","Low hallucination rate"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1000000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":1.25,"output":2.5,"cachedInput":0.2,"note":"Prompts of 200K+ tokens: $2.50 input / $5 output per 1M tokens"},"benchmarks":{},"benchmarksAsOf":{},"popularity":55,"links":{"website":"https://x.ai"},"sources":["https://docs.x.ai/docs/models"],"providerName":"xAI (SpaceXAI)","blendedPrice":1.563},{"id":"gemini-embedding-2","name":"Gemini Embedding 2","providerId":"google","family":"Gemini Embedding","kind":"embedding","releaseDate":"2026-05-20","status":"ga","summary":"Google’s first multimodal embedding model.","description":"Gemini Embedding 2 is Google’s first multimodal embedding model: it places text and images in one shared vector space, so a text query can find images and vice versa for search and RAG.","useCases":["multimodal","research","enterprise"],"strengths":["Text + image in one space","Available on Gemini API and Vertex AI"],"technologies":["vector-search","rag","native-multimodal"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":55,"links":{"website":"https://ai.google.dev/gemini-api/docs/embeddings"},"sources":[],"providerName":"Google","blendedPrice":null},{"id":"minimax-m3","name":"MiniMax M3","providerId":"minimax","family":"MiniMax M","kind":"llm","releaseDate":"2026-05-31","status":"ga","summary":"Open multimodal model built for long-horizon agentic work.","description":"MiniMax M3 is an MIT-licensed mixture-of-experts model from MiniMax that reads text, images and video. MiniMax built it for long agent tasks and coding.","useCases":["agents","coding","long-context","multimodal"],"strengths":["Open weights (MIT)","Long agent tasks"],"technologies":["reasoning","moe"],"openWeights":true,"license":"MIT","huggingFaceId":"MiniMaxAI/Minimax-M3","architecture":{"type":"MoE","totalParams":"428B"},"contextWindow":1000000,"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":0.3,"output":1.2,"cachedInput":0.06,"note":"Prompts over 512K tokens: $0.60 input / $2.40 output per 1M tokens"},"usage":{"tokenShare":0.67,"usageRank":24,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":40.6,"codingScore":30.9,"agentScore":26,"mathScore":25.1,"toolCallingScore":19.8,"visionScore":25.7,"sweBenchPro":59,"gpqa":90.4,"tauBenchAirline":75,"arenaText":1439,"arenaWebdev":1483,"eci":146.9},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":54,"links":{"huggingFace":"https://huggingface.co/MiniMaxAI/Minimax-M3"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://platform.minimax.io/docs/pricing/overview","https://platform.minimax.io/docs/guides/text-generation","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"MiniMax","blendedPrice":0.525},{"id":"qwen3.8-2.4t-a95b","name":"Qwen3.8 2.4T A95B","providerId":"alibaba","family":"Qwen3.8","kind":"llm","releaseDate":"2026-08-12","status":"ga","summary":"The open-weight release of Qwen3.8 Max — one of the largest open models.","description":"Qwen3.8 2.4T A95B is the open-weight counterpart of Qwen3.8 Max: a sparse mixture of experts that activates 95B of its 2.4 trillion parameters per token.","useCases":["reasoning","coding","on-device","enterprise"],"strengths":["Frontier-scale open weights","Self-hostable for data sovereignty"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Qwen License","huggingFaceId":"Qwen/Qwen3.8-2.4T-A95B","architecture":{"type":"MoE","totalParams":"2.4T","activeParams":"95B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":2,"output":6},"usage":{"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"gpqa":87.1,"tauBenchAirline":75.3},"benchmarksAsOf":{"gpqa":"2026-10-30","tauBenchAirline":"2026-10-30"},"popularity":52,"links":{"huggingFace":"https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"},"sources":["https://www.alibabacloud.com/help/en/model-studio/model-pricing","https://openrouter.ai/rankings"],"providerName":"Alibaba (Qwen)","blendedPrice":3},{"id":"voyage-4-large","name":"voyage-4-large","providerId":"voyage","family":"Voyage 4","kind":"embedding","releaseDate":"2026-07-27","status":"ga","summary":"Voyage’s highest-quality general and multilingual retrieval embedding.","description":"voyage-4-large is Voyage AI’s top general-purpose, multilingual embedding model for retrieval. Voyage trained it so vectors can be shortened to 1024, 512 or 256 dimensions from the full 2048 and stored in several quantised formats.","useCases":["research","long-context","enterprise"],"strengths":["Matryoshka dimensions (256–2048)","Quantized embeddings"],"technologies":["vector-search","rag","quantization"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":52,"links":{"website":"https://www.voyageai.com","docs":"https://docs.voyageai.com/docs/embeddings"},"sources":[],"providerName":"Voyage AI (MongoDB)","blendedPrice":null},{"id":"mistral-large-4","name":"Mistral Large 4","providerId":"mistral","family":"Mistral Large","kind":"llm","releaseDate":"2026-10-06","status":"preview","summary":"Mistral’s 1T-parameter multimodal MoE flagship (preview; weights promised).","description":"Mistral Large 4 is Mistral’s new flagship, a text-and-image mixture of experts with 1T total and 52B active parameters, built for reasoning, coding and agents. Mistral has promised open weights but not yet released them.","useCases":["reasoning","coding","agents","enterprise","long-context"],"strengths":["European provider (data residency)","Weights promised"],"technologies":["reasoning","moe"],"openWeights":false,"architecture":{"type":"MoE","totalParams":"1T","activeParams":"52B"},"contextWindow":1000000,"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":0.68,"output":2.09,"cachedInput":0.07,"note":"Discounted price; list price $1.36 input / $4.18 output per 1M tokens"},"usage":{"tokenShare":0.02,"usageRank":65,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":45.2,"codingScore":36.8,"agentScore":35,"mathScore":35.5,"toolCallingScore":30.9,"visionScore":1.4,"arenaText":1432,"arenaWebdev":1544,"eci":150.6},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":50,"links":{"website":"https://mistral.ai","docs":"https://docs.mistral.ai"},"sources":["https://docs.mistral.ai/models/mistral-large-4-0","https://mistral.ai/news/mistral-large-4","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Mistral AI","blendedPrice":1.033},{"id":"gemma-4-26b-a4b","name":"Gemma 4 26B A4B","providerId":"google","family":"Gemma 4","kind":"llm","releaseDate":"2026-04-03","knowledgeCutoff":"2025-01-01","status":"ga","summary":"Open MoE Gemma with only 3.8B active parameters.","description":"Gemma 4 26B A4B is Google DeepMind’s mixture-of-experts Gemma 4 model. Only 3.8B of its 25.2B parameters run per token, so it is much cheaper to serve than the dense 31B model. It reads text, images and video.","useCases":["on-device","vision","chat","high-volume"],"strengths":["3.8B active parameters","Open weights"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Gemma Terms of Use","huggingFaceId":"google/gemma-4-26B-A4B-it","architecture":{"type":"MoE","totalParams":"25.2B","activeParams":"3.8B"},"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.14,"usageRank":51,"asOf":"2026-10-31","topTasks":[{"label":"Summarization","share":6.1},{"label":"Memory Extraction","share":5.5},{"label":"Translation","share":3.2}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":28.6,"agentScore":12.3,"mathScore":25.2,"toolCallingScore":12.3,"visionScore":14.6,"gpqa":70.5,"tauBenchAirline":68.7,"arenaText":1438,"arenaWebdev":1359,"eci":142},"benchmarksAsOf":{"reasoningScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":50,"links":{"website":"https://deepmind.google/models/gemma/","huggingFace":"https://huggingface.co/google/gemma-4-26B-A4B-it"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Google","blendedPrice":null},{"id":"inkling","name":"Inkling","providerId":"thinking-machines","family":"Inkling","kind":"llm","releaseDate":"2026-07-17","status":"ga","summary":"Thinking Machines’ first open-weight frontier MoE.","description":"Inkling is an Apache-licensed mixture of experts from Thinking Machines Lab, with 975B total and 41B active parameters. It covers general reasoning, coding, agents and tool use, and accepts text, images and audio.","useCases":["reasoning","coding","agents","multimodal"],"strengths":["Apache 2.0 weights","Audio input"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"thinkingmachines/Inkling","architecture":{"type":"MoE","totalParams":"975B","activeParams":"41B"},"modalities":{"input":["text","image","audio"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.04,"usageRank":61,"asOf":"2026-10-31","topTasks":[{"label":"DevOps & Config","share":2.3}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":36,"codingScore":26.3,"agentScore":19.5,"mathScore":28.5,"toolCallingScore":16.2,"visionScore":15.6,"gpqa":83.4,"tauBenchAirline":67.2,"arenaText":1442,"arenaWebdev":1415,"eci":148.5},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":50,"links":{"website":"https://thinkingmachines.ai","huggingFace":"https://huggingface.co/thinkingmachines/Inkling"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Thinking Machines Lab","blendedPrice":null},{"id":"kling-3","name":"Kling 3.0","providerId":"kuaishou","family":"Kling","kind":"video","releaseDate":"2026-02-04","status":"ga","summary":"Kuaishou’s 1080p video model, popular with creators.","description":"Kling 3.0 is Kuaishou’s video generation model, offered in 1080p Pro and Omni variants through the Kling app and API.","useCases":["multimodal"],"strengths":["1080p output","Large creator community"],"technologies":["diffusion"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["video"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":50,"links":{"website":"https://klingai.com"},"sources":[],"providerName":"Kuaishou (Kling)","blendedPrice":null},{"id":"mimo-v2.6-flash","name":"MiMo-V2.6-Flash","providerId":"xiaomi","family":"MiMo","kind":"llm","releaseDate":"2026-09-21","status":"ga","summary":"Open 309B MoE with omni-modal input.","description":"MiMo-V2.6-Flash is Xiaomi’s smaller open mixture of experts (309B total, 15B active), using hybrid attention to handle long inputs efficiently. It accepts text, images, audio and video.","useCases":["high-volume","multimodal","on-device","extraction"],"strengths":["Audio & video input","Open weights (MIT)"],"technologies":["reasoning","moe"],"openWeights":true,"license":"MIT","huggingFaceId":"XiaomiMiMo/MiMo-V2.6-Flash-RL","architecture":{"type":"MoE","totalParams":"309B","activeParams":"15B"},"modalities":{"input":["text","image","audio","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":4.59,"usageRank":4,"asOf":"2026-10-31","topTasks":[{"label":"Frontend & UI","share":5.8},{"label":"SQL & Database","share":3.3},{"label":"DevOps","share":3.1}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":42.8,"codingScore":34.4,"agentScore":32.2,"toolCallingScore":26.7,"visionScore":26.1,"gpqa":74.2,"tauBenchAirline":78.2,"arenaText":1448,"arenaWebdev":1637},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28"},"popularity":48,"links":{"huggingFace":"https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Xiaomi","blendedPrice":null},{"id":"kimi-k2.7-code","name":"Kimi K2.7 Code","providerId":"moonshot","family":"Kimi K","kind":"llm","releaseDate":"2026-06-12","status":"ga","summary":"Coding-focused Kimi K2 model for long, end-to-end programming tasks.","description":"Kimi K2.7 Code is a coding specialist in Moonshot AI’s Kimi K2 family. It is a multimodal mixture of experts designed to carry programming tasks through to completion over long sessions.","useCases":["coding","agents"],"strengths":["Open weights","Long-horizon coding"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Modified MIT","huggingFaceId":"moonshotai/Kimi-K2.7-Code","architecture":{"type":"MoE"},"contextWindow":262144,"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":0.95,"output":4,"cachedInput":0.19},"usage":{"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":33.4,"codingScore":29.2,"agentScore":26.2,"mathScore":18.8,"toolCallingScore":23.2,"gpqa":86.2,"tauBenchAirline":73.8,"arenaWebdev":1474,"eci":150.2},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":48,"links":{"website":"https://www.moonshot.ai","huggingFace":"https://huggingface.co/moonshotai/Kimi-K2.7-Code"},"sources":["https://platform.kimi.ai/docs/pricing/chat","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Moonshot AI","blendedPrice":1.713},{"id":"nova-2-lite","name":"Nova 2 Lite","providerId":"amazon","family":"Nova 2","kind":"llm","releaseDate":"2025-12-02","status":"ga","summary":"Amazon’s low-cost multimodal model, native to AWS Bedrock.","description":"Amazon Nova 2 Lite is a fast, low-cost multimodal model accepting text, image, video and documents, available through AWS Bedrock.","useCases":["enterprise","high-volume","multimodal","long-context"],"strengths":["AWS-native","Video & document input"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image","video","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":null,"output":null},"benchmarks":{"reasoningScore":26.3,"codingScore":12.1,"agentScore":4.4,"mathScore":24.3,"toolCallingScore":10.8,"visionScore":5.1,"arenaText":1335},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","arenaText":"2026-10-28"},"popularity":46,"links":{"website":"https://aws.amazon.com/ai/generative-ai/nova/"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Amazon","blendedPrice":null},{"id":"wan-3","name":"Wan 3.0","providerId":"alibaba","family":"Wan","kind":"video","releaseDate":"2026-08-19","status":"ga","summary":"Alibaba’s latest Wan video generation model.","description":"Wan 3.0 is the newest model in Alibaba’s Wan video generation family, creating video from text or a starting image.","useCases":["multimodal","image-generation"],"strengths":["Text- or image-to-video"],"technologies":["diffusion"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["video"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToVideo":1476,"arenaImageToVideo":1480},"benchmarksAsOf":{"arenaTextToVideo":"2026-09-22","arenaImageToVideo":"2026-09-22"},"popularity":46,"links":{"website":"https://wan.video"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Alibaba (Qwen)","blendedPrice":null},{"id":"genie-3","name":"Genie 3","providerId":"google","family":"Genie","kind":"world-model","releaseDate":"2025-08-05","status":"limited","summary":"Real-time interactive world model that generates explorable 3D environments.","description":"Genie 3 is Google DeepMind’s general-purpose world model: it turns a text description into a photorealistic environment you can explore in real time at 720p and 20–24 fps, with “promptable world events” that change weather or add objects mid-session. Worlds stay consistent for a few minutes. Available through the experimental Project Genie prototype.","useCases":["multimodal","research","agents"],"strengths":["Interactive in real time","Promptable world events","Minutes-long consistency"],"technologies":["native-multimodal"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["video"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":46,"links":{"website":"https://deepmind.google/models/genie/"},"sources":["https://deepmind.google/models/genie/","https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/"],"providerName":"Google","blendedPrice":null},{"id":"minimax-m3.1-flash","name":"MiniMax M3.1 Flash","providerId":"minimax","family":"MiniMax M","kind":"llm","releaseDate":"2026-10-19","status":"ga","summary":"Low-cost, sparser MiniMax model for agents, coding and computer use.","description":"MiniMax M3.1 Flash is a cheaper, sparser sibling of MiniMax M3 aimed at agent workflows, coding, computer use and security tasks. It reads text, images and video with a 1M-token context. MiniMax says the weights will be published later.","useCases":["agents","coding","high-volume","multimodal"],"strengths":["Low price","Computer use","1M-token context"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"contextWindow":1000000,"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":0.15,"output":0.6,"cachedInput":0.015},"usage":{"tokenShare":2.47,"usageRank":12,"asOf":"2026-10-31","topTasks":[{"label":"Repo Scanning","share":3.2},{"label":"DevOps & Config","share":2.5},{"label":"DevOps","share":2.4}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":45.2,"codingScore":38.7,"agentScore":36,"mathScore":24.3,"toolCallingScore":27,"visionScore":28.8,"gpqa":86.7,"tauBenchAirline":66,"arenaWebdev":1582},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaWebdev":"2026-10-28"},"popularity":45,"links":{},"sources":["https://www.minimax.io/blog/minimax-m31f","https://platform.minimax.io/docs/pricing/overview","https://platform.minimax.io/docs/guides/text-generation","https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"MiniMax","blendedPrice":0.262},{"id":"mistral-medium-3.5","name":"Mistral Medium 3.5","providerId":"mistral","family":"Mistral Medium","kind":"llm","releaseDate":"2026-04-30","status":"ga","summary":"Enterprise-grade multimodal model from Mistral.","description":"Mistral Medium 3.5 is Mistral’s mid-size model for enterprise deployments, accepting text, images and documents.","useCases":["enterprise","chat","extraction","realtime"],"strengths":["On-prem deployment options","Document input"],"technologies":["reasoning"],"openWeights":false,"architecture":{"totalParams":"128B"},"contextWindow":256000,"modalities":{"input":["text","image","file"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":true},"pricing":{"input":1.5,"output":7.5},"benchmarks":{"reasoningScore":26.2,"codingScore":23.7,"agentScore":7.5,"mathScore":18.9,"toolCallingScore":2,"arenaText":1426,"arenaWebdev":1264,"eci":141.5},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":45,"links":{"docs":"https://docs.mistral.ai"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://docs.mistral.ai/models/mistral-medium-3-5-26-04","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Mistral AI","blendedPrice":3},{"id":"beam-1.0","name":"Beam 1.0","providerId":"reflection","family":"Beam","kind":"llm","releaseDate":"2026-10-16","knowledgeCutoff":"2026-06-30","status":"ga","summary":"Reflection AI’s first open-weight model, a 501B MoE for coding and agents.","description":"Beam 1.0 is a sparse mixture of experts (501B total, 23B active) from Reflection AI, released under Apache 2.0 for advanced coding, reasoning and long agent tasks. It reads and writes text, mainly English, with a 1M-token context and MCP-based tool use. BF16, FP8 and NVFP4 checkpoints are on Hugging Face; Reflection also runs a hosted API.","useCases":["coding","agents","reasoning","long-context"],"strengths":["Open weights (Apache 2.0)","1M-token context","Only 23B active parameters"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"reflection/Beam-1.0-501B-A23B","architecture":{"type":"MoE","totalParams":"501B","activeParams":"23B"},"contextWindow":1000000,"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"reasoningScore":45.1,"codingScore":33.8,"agentScore":28.8,"mathScore":35.1,"toolCallingScore":26.9,"visionScore":16.1,"sweBenchPro":66.1,"arenaText":1436,"arenaWebdev":1436},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","arenaText":"2026-10-28","arenaWebdev":"2026-10-28"},"popularity":45,"links":{"website":"https://reflection.ai","huggingFace":"https://huggingface.co/reflection/Beam-1.0-501B-A23B"},"sources":["https://reflection.ai/blog/beam-is-available","https://llm-stats.com/leaderboards/llm-leaderboard","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Reflection AI","blendedPrice":null},{"id":"nemotron-3-super","name":"Nemotron 3 Super","providerId":"nvidia","family":"Nemotron 3","kind":"llm","releaseDate":"2026-03-11","knowledgeCutoff":"2025-06-01","status":"ga","summary":"Open hybrid Mamba-Transformer MoE for multi-agent applications.","description":"NVIDIA Nemotron 3 Super is an open hybrid mixture of experts with 120B parameters, 12B of them active, designed to stay compute-efficient in complex multi-agent systems.","useCases":["agents","on-device","enterprise","realtime"],"strengths":["Hybrid Mamba-Transformer","Optimized for NVIDIA GPUs"],"technologies":["hybrid-ssm","reasoning","moe"],"openWeights":true,"license":"NVIDIA Open Model License","huggingFaceId":"nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8","architecture":{"type":"Hybrid Mamba-Transformer MoE","totalParams":"120B","activeParams":"12B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.22,"usageRank":47,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":24.4,"codingScore":13,"agentScore":-1.1,"mathScore":29.9,"toolCallingScore":2.4,"visionScore":9,"arenaText":1360},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","arenaText":"2026-10-28"},"popularity":44,"links":{"huggingFace":"https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"NVIDIA","blendedPrice":null},{"id":"grok-imagine-image-2","name":"Grok Imagine Image 2.0","providerId":"xai","family":"Grok Imagine","kind":"image","releaseDate":"2026-08-07","status":"ga","summary":"SpaceXAI’s text-to-image model, in Grok and via API.","description":"Grok Imagine Image 2.0 is SpaceXAI’s (formerly xAI) text-to-image model, available in Grok and via API.","useCases":["image-generation"],"strengths":["Available in Grok and via API"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToImage":1317,"arenaImageEdit":1439},"benchmarksAsOf":{"arenaTextToImage":"2026-10-26","arenaImageEdit":"2026-10-27"},"popularity":44,"links":{"website":"https://x.ai"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"xAI (SpaceXAI)","blendedPrice":null},{"id":"seedance-2.5","name":"Dreamina Seedance 2.5","providerId":"bytedance","family":"Seedance","kind":"video","releaseDate":"2026-07-31","status":"ga","summary":"ByteDance Seed’s flagship video model.","description":"Dreamina Seedance 2.5 is ByteDance Seed’s flagship video generation model, available in Dreamina and via API.","useCases":["multimodal"],"strengths":["Available in Dreamina and via API"],"technologies":["diffusion"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["video"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToVideo":1474,"arenaImageToVideo":1477},"benchmarksAsOf":{"arenaTextToVideo":"2026-09-22","arenaImageToVideo":"2026-09-22"},"popularity":44,"links":{"website":"https://seed.bytedance.com"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"ByteDance Seed","blendedPrice":null},{"id":"scribe-v2","name":"Scribe v2","providerId":"elevenlabs","family":"Scribe","kind":"transcription","releaseDate":"2026-01-09","status":"ga","summary":"ElevenLabs’ highly accurate transcription model.","description":"Scribe v2 is ElevenLabs’ speech-to-text model, with speaker diarization.","useCases":["extraction","multimodal"],"strengths":["Speaker diarization"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["audio"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":44,"links":{"website":"https://elevenlabs.io/speech-to-text"},"sources":[],"providerName":"ElevenLabs","blendedPrice":null},{"id":"llama-guard-4","name":"Llama Guard 4 12B","providerId":"meta","family":"Llama Guard","kind":"safety","releaseDate":"2025-04-29","status":"ga","summary":"Open multimodal safety classifier for prompts and responses.","description":"Llama Guard 4 is derived from Llama 4 Scout and fine-tuned for content-safety classification. It classifies both prompts (input) and responses (output) — including images — against a hazard taxonomy, returning safe/unsafe plus the violated categories.","useCases":["enterprise","extraction","vision"],"strengths":["Open weights","Text + image moderation","Standard hazard taxonomy"],"technologies":["guardrails"],"openWeights":true,"license":"Llama 4 Community License","huggingFaceId":"meta-llama/Llama-Guard-4-12B","architecture":{"totalParams":"12B"},"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":44,"links":{"website":"https://www.llama.com/docs/model-cards-and-prompt-formats/llama-guard-4/","huggingFace":"https://huggingface.co/meta-llama/Llama-Guard-4-12B"},"sources":[],"providerName":"Meta","blendedPrice":null},{"id":"command-a-plus","name":"Command A+","providerId":"cohere","family":"Command","kind":"llm","releaseDate":"2026-05-20","status":"ga","summary":"Cohere’s enterprise agent model with strict tool schemas.","description":"Command A+ is Cohere’s flagship for enterprise agent workflows. It reads text and images and supports native tool calling with strict tool schemas and structured outputs.","useCases":["enterprise","agents","extraction","realtime"],"strengths":["Strict tool schemas","Private deployment"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"CohereLabs/command-a-plus-05-2026-bf16","architecture":{"type":"MoE","totalParams":"218B"},"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"reasoningScore":21.1,"codingScore":7.5,"mathScore":23.8,"toolCallingScore":10,"visionScore":12.2},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31"},"popularity":42,"links":{"website":"https://cohere.com","huggingFace":"https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://cohere.com/blog/command-a-plus","https://docs.cohere.com/docs/models"],"providerName":"Cohere","blendedPrice":null},{"id":"mercury-3","name":"Mercury 3","providerId":"inception","family":"Mercury","kind":"llm","releaseDate":"2026-10-19","status":"ga","summary":"Inception’s flagship diffusion LLM, with a 1M-token context.","description":"Mercury 3 is Inception’s flagship diffusion language model. Like earlier Mercury models it refines many tokens in parallel instead of writing them one by one; Inception reports over 1,000 tokens per second on NVIDIA GPUs (vendor-reported). It raises the context window to 1M tokens and offers four reasoning-effort levels, from instant to high.","useCases":["realtime","high-volume","coding","agents","long-context"],"strengths":["Parallel (diffusion) generation","1M-token context","Low price"],"technologies":["reasoning","diffusion-lm"],"openWeights":false,"architecture":{"type":"Diffusion LLM"},"contextWindow":1000000,"maxOutputTokens":131072,"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":0.1,"output":0.35,"cachedInput":0.01,"note":"Launch price (50% off, end date not announced); standard price $0.20 input / $0.02 cached / $0.70 output per 1M tokens"},"benchmarks":{"reasoningScore":34.6,"codingScore":21.7,"agentScore":19.3,"toolCallingScore":15.1},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","toolCallingScore":"2026-10-31"},"popularity":42,"links":{"website":"https://www.inceptionlabs.ai"},"sources":["https://docs.inceptionlabs.ai/get-started/models","https://www.inceptionlabs.ai/blog/introducing-mercury-3","https://llm-stats.com/leaderboards/llm-leaderboard"],"providerName":"Inception","blendedPrice":0.163},{"id":"seedream-5-pro","name":"Seedream 5.0 Pro","providerId":"bytedance","family":"Seedream","kind":"image","releaseDate":"2026-07-08","status":"ga","summary":"ByteDance Seed’s flagship image model.","description":"Seedream 5.0 Pro is ByteDance Seed’s flagship text-to-image model.","useCases":["image-generation"],"strengths":["ByteDance Seed flagship","Reference-image input"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToImage":1250,"arenaImageEdit":1390},"benchmarksAsOf":{"arenaTextToImage":"2026-10-26","arenaImageEdit":"2026-10-27"},"popularity":42,"links":{"website":"https://seed.bytedance.com"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"ByteDance Seed","blendedPrice":null},{"id":"gemini-omni-flash-1.1","name":"Gemini Omni Flash 1.1","providerId":"google","family":"Gemini Omni","kind":"video","releaseDate":"2026-08-28","status":"ga","summary":"Google’s fast video generation model.","description":"Gemini Omni Flash 1.1 is a Google video generation model that creates clips from text or a starting image.","useCases":["multimodal"],"strengths":["Text- or image-to-video"],"technologies":["native-multimodal"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["video"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToVideo":1516,"arenaImageToVideo":1488},"benchmarksAsOf":{"arenaTextToVideo":"2026-09-22","arenaImageToVideo":"2026-09-22"},"popularity":42,"links":{"website":"https://deepmind.google"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Google","blendedPrice":null},{"id":"nemotron-3.5-lightning","name":"Nemotron 3.5 Lightning","providerId":"nvidia","family":"Nemotron 3.5","kind":"llm","releaseDate":"2026-08-11","status":"ga","summary":"Small open 30B-A3B model tuned for low latency.","description":"Nemotron 3.5 Lightning is an open NVIDIA model with 30B total and 3B active parameters, tuned for quick responses and high throughput.","useCases":["realtime","high-volume","on-device","extraction"],"strengths":["3B active parameters","Low-latency design (vendor-reported)"],"technologies":["reasoning","moe"],"openWeights":true,"license":"NVIDIA Open Model License","huggingFaceId":"nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16","architecture":{"type":"MoE","totalParams":"30B","activeParams":"3B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.52,"usageRank":28,"asOf":"2026-10-31","topTasks":[{"label":"DevOps","share":4.6},{"label":"Math","share":3}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":19.4,"codingScore":7.8,"agentScore":4.5,"mathScore":17.6,"toolCallingScore":-1.7,"visionScore":-0.5,"gpqa":70.3,"tauBenchAirline":65.7,"arenaText":1353},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28"},"popularity":40,"links":{"huggingFace":"https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"NVIDIA","blendedPrice":null},{"id":"nemotron-3.5-super-vl","name":"Nemotron 3.5 Super VL","providerId":"nvidia","family":"Nemotron 3.5","kind":"llm","releaseDate":"2026-10-15","status":"ga","summary":"Open multimodal MoE for documents, GUIs and video.","description":"Nemotron 3.5 Super VL is NVIDIA’s open vision-language mixture of experts (120B total, 12B active). It reads text, images, documents, screenshots of user interfaces and video.","useCases":["vision","agents","extraction","long-context"],"strengths":["GUI and document understanding","Open weights"],"technologies":["reasoning","moe"],"openWeights":true,"huggingFaceId":"nvidia/NVIDIA-Nemotron-3.5-Super-VL","architecture":{"type":"MoE","totalParams":"120B","activeParams":"12B"},"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.48,"usageRank":29,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":35.7,"codingScore":21.3,"agentScore":16.4,"mathScore":27.9,"toolCallingScore":11,"visionScore":20.3,"sweBenchPro":51.8,"gpqa":87.4,"tauBenchAirline":73.3,"arenaText":1412,"arenaWebdev":1399},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28"},"popularity":40,"links":{"huggingFace":"https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"NVIDIA","blendedPrice":null},{"id":"nemotron-3-ultra","name":"Nemotron 3 Ultra","providerId":"nvidia","family":"Nemotron 3","kind":"llm","releaseDate":"2026-06-04","knowledgeCutoff":"2025-09-30","status":"ga","summary":"NVIDIA’s largest open model: hybrid Mamba-Transformer MoE, 55B active.","description":"Nemotron 3 Ultra is NVIDIA’s largest open model, aimed at frontier reasoning and orchestrating other agents. It is a hybrid Transformer-Mamba mixture of experts with 55B of 550B parameters active.","useCases":["reasoning","agents","enterprise","on-device"],"strengths":["Open weights","Hybrid Mamba for long inputs","Orchestration"],"technologies":["hybrid-ssm","reasoning","moe"],"openWeights":true,"license":"NVIDIA Open Model License","huggingFaceId":"nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16","architecture":{"type":"Hybrid Mamba-Transformer MoE","totalParams":"550B","activeParams":"55B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":3.03,"usageRank":10,"asOf":"2026-10-31","topTasks":[{"label":"DevOps","share":4.6},{"label":"DevOps & Config","share":3.4},{"label":"Shell Execution","share":2.6}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":35,"codingScore":20,"agentScore":11.9,"mathScore":36,"toolCallingScore":1.8,"visionScore":17.7,"gpqa":81.2,"tauBenchAirline":72,"arenaText":1427,"eci":146},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","eci":"2026-10-31"},"popularity":40,"links":{"website":"https://www.nvidia.com/en-us/ai/","huggingFace":"https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"NVIDIA","blendedPrice":null},{"id":"inkling-small","name":"Inkling Small","providerId":"thinking-machines","family":"Inkling","kind":"llm","releaseDate":"2026-07-30","status":"ga","summary":"The smaller, cheaper-to-serve member of the Inkling family.","description":"Inkling Small is the efficient model in Thinking Machines Lab’s open Inkling family: a multimodal mixture of experts with 12B of 276B parameters active.","useCases":["reasoning","coding","agents","high-volume"],"strengths":["Apache 2.0 weights","12B active parameters"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"thinkingmachines/Inkling-Small","architecture":{"type":"MoE","totalParams":"276B","activeParams":"12B"},"modalities":{"input":["text","image","audio"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":37,"codingScore":28.6,"agentScore":22.4,"mathScore":32.6,"toolCallingScore":18.5,"visionScore":20.2,"sweBenchPro":55.9,"gpqa":86.6,"tauBenchAirline":60.7,"arenaText":1406,"arenaWebdev":1409,"eci":150.5},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28","arenaWebdev":"2026-10-28","eci":"2026-10-31"},"popularity":40,"links":{"website":"https://thinkingmachines.ai","huggingFace":"https://huggingface.co/thinkingmachines/Inkling-Small"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","https://epoch.ai/benchmarks"],"providerName":"Thinking Machines Lab","blendedPrice":null},{"id":"hy3","name":"Hy3","providerId":"tencent","family":"Hy","kind":"llm","releaseDate":"2026-07-06","status":"ga","summary":"Tencent’s open 295B MoE with configurable reasoning effort.","description":"Hy3 is an Apache-licensed Tencent mixture of experts with 295B parameters, 21B active, using 192 experts with top-8 routing. Tencent built it for reasoning, agents and production use, with adjustable reasoning effort.","useCases":["reasoning","agents","high-volume"],"strengths":["Apache 2.0 weights","Adjustable reasoning effort"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"tencent/Hy3","architecture":{"type":"MoE","totalParams":"295B","activeParams":"21B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":1.55,"usageRank":14,"asOf":"2026-10-31","topTasks":[{"label":"DevOps","share":4.2},{"label":"Frontend & UI","share":3.9},{"label":"File I/O","share":3.5}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":42.3,"codingScore":32.1,"agentScore":26.9,"mathScore":35.2,"toolCallingScore":19.7,"sweBenchPro":57.9,"arenaText":1454,"arenaWebdev":1505},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","sweBenchPro":"2026-10-31","arenaText":"2026-10-28","arenaWebdev":"2026-10-28"},"popularity":40,"links":{"huggingFace":"https://huggingface.co/tencent/Hy3"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Tencent","blendedPrice":null},{"id":"voyage-code-4","name":"voyage-code-4","providerId":"voyage","family":"Voyage 4","kind":"embedding","releaseDate":"2026-08-13","status":"ga","summary":"Code-retrieval embedding built for coding agents.","description":"voyage-code-4 is Voyage AI’s embedding model for source code, made for code search and coding agents. Voyage AI is part of MongoDB. Its Matryoshka vectors can be cut to 2048, 1024, 512 or 256 dimensions.","useCases":["coding","agents"],"strengths":["Tuned for code search","Matryoshka dimensions"],"technologies":["vector-search","rag"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":40,"links":{"website":"https://www.voyageai.com","docs":"https://docs.voyageai.com/docs/embeddings"},"sources":[],"providerName":"Voyage AI (MongoDB)","blendedPrice":null},{"id":"mai-image-2.6","name":"MAI-Image-2.6","providerId":"microsoft","family":"MAI-Image","kind":"image","releaseDate":"2026-08-10","status":"ga","summary":"Microsoft AI’s in-house image model.","description":"MAI-Image-2.6 is Microsoft AI’s in-house text-to-image model, with a cheaper Flash variant.","useCases":["image-generation"],"strengths":["In-house Microsoft model","Cheaper Flash variant"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToImage":1329,"arenaImageEdit":1427},"benchmarksAsOf":{"arenaTextToImage":"2026-10-26","arenaImageEdit":"2026-10-27"},"popularity":40,"links":{"website":"https://microsoft.ai"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Microsoft AI","blendedPrice":null},{"id":"qwen-image-3-pro","name":"Qwen-Image-3.0-Pro","providerId":"alibaba","family":"Qwen-Image","kind":"image","releaseDate":"2026-07-21","status":"ga","summary":"Alibaba’s flagship image model with strong text rendering.","description":"Qwen-Image-3.0-Pro is the top tier of Alibaba’s Qwen-Image family of text-to-image models.","useCases":["image-generation"],"strengths":["Top tier of Qwen-Image 3.0"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToImage":1254},"benchmarksAsOf":{"arenaTextToImage":"2026-10-26"},"popularity":40,"links":{"website":"https://qwen.ai"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Alibaba (Qwen)","blendedPrice":null},{"id":"minimax-h3","name":"MiniMax H3","providerId":"minimax","family":"Hailuo","kind":"video","releaseDate":"2026-07-30","status":"ga","summary":"MiniMax’s open-weight video generation model.","description":"MiniMax H3 (Hailuo) is MiniMax’s video generation model, released with open weights and also offered through MiniMax’s API.","useCases":["multimodal","on-device"],"strengths":["Open weights","Text- or image-to-video"],"technologies":["diffusion"],"openWeights":true,"huggingFaceId":"MiniMaxAI/MiniMax-H3","architecture":{},"modalities":{"input":["text","image"],"output":["video"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToVideo":1460,"arenaImageToVideo":1495},"benchmarksAsOf":{"arenaTextToVideo":"2026-09-22","arenaImageToVideo":"2026-09-22"},"popularity":40,"links":{"website":"https://hailuoai.video","huggingFace":"https://huggingface.co/MiniMaxAI/MiniMax-H3"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"MiniMax","blendedPrice":null},{"id":"runway-gen-4.5","name":"Runway Gen-4.5","providerId":"runway","family":"Gen","kind":"video","releaseDate":"2025-12-01","status":"ga","summary":"Runway’s cinematic text- and image-to-video model.","description":"Runway Gen-4.5 is Runway’s video model for text-to-video and image-to-video, designed for cinematic scenes with convincing motion and close prompt adherence.","useCases":["multimodal"],"strengths":["Cinematic motion","Mature creative tooling"],"technologies":["diffusion"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["video"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToVideo":1225},"benchmarksAsOf":{"arenaTextToVideo":"2026-09-22"},"popularity":40,"links":{"website":"https://runwayml.com"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Runway","blendedPrice":null},{"id":"gemini-3.8-flash-tts","name":"Gemini 3.8 Flash TTS","providerId":"google","family":"Gemini TTS","kind":"speech","releaseDate":"2026-09-23","status":"ga","summary":"Google’s controllable TTS with a cheaper Flash-Lite tier.","description":"Gemini 3.8 Flash TTS is Google’s text-to-speech model in the Gemini API, with a lower-cost Flash-Lite TTS variant.","useCases":["multimodal","high-volume"],"strengths":["Controllable speech","Cheaper Flash-Lite tier"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["audio"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":40,"links":{"website":"https://ai.google.dev/gemini-api/docs/speech-generation"},"sources":[],"providerName":"Google","blendedPrice":null},{"id":"universal-3-pro","name":"Universal-3 Pro","providerId":"assemblyai","family":"Universal","kind":"transcription","releaseDate":"2026-02-04","status":"ga","summary":"AssemblyAI’s flagship speech-to-text model.","description":"Universal-3 Pro is AssemblyAI’s top speech-to-text model, available through its API alongside speech-understanding features.","useCases":["extraction","enterprise"],"strengths":["Speech-understanding features","Mature developer API"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["audio"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":40,"links":{"website":"https://www.assemblyai.com"},"sources":[],"providerName":"AssemblyAI","blendedPrice":null},{"id":"gpt-transcribe","name":"GPT Transcribe","providerId":"openai","family":"GPT Transcribe","kind":"transcription","releaseDate":"2026-08-05","status":"ga","summary":"OpenAI’s speech-to-text model with context and keyword hints.","description":"GPT Transcribe is OpenAI’s speech-to-text model for recorded audio, streamed files and Realtime turns. You can give it free-form context and keyword hints, and it handles many languages.","useCases":["extraction","realtime","enterprise"],"strengths":["Keyword hints","Free-form context"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["audio"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":40,"links":{},"sources":[],"providerName":"OpenAI","blendedPrice":null},{"id":"cosmos-3","name":"Cosmos 3","providerId":"nvidia","family":"Cosmos","kind":"world-model","releaseDate":"2026-06-01","status":"ga","summary":"NVIDIA’s open omni-modal world foundation model for physical AI.","description":"Cosmos 3 is NVIDIA’s open world foundation model for robotics, autonomous vehicles and vision agents. A mixture-of-transformers design pairs a reasoning transformer with a generation transformer so one model can understand and generate text, images, video, ambient sound and actions — for simulation, synthetic data and as a backbone for robot policies. Ships as Super (highest fidelity) and Nano (fast) tiers, with Edge to follow.","useCases":["multimodal","agents","research"],"strengths":["Open weights on Hugging Face","Understands and generates actions","NIM deployment"],"technologies":["native-multimodal","diffusion"],"openWeights":true,"architecture":{"type":"Mixture of transformers"},"modalities":{"input":["text","image","video","audio"],"output":["text","video","action"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":40,"links":{"website":"https://www.nvidia.com/en-us/ai/cosmos/"},"sources":["https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Cosmos-3-the-Open-Frontier-Foundation-Model-for-Physical-AI/default.aspx"],"providerName":"NVIDIA","blendedPrice":null},{"id":"mercury-2.5","name":"Mercury 2.5","providerId":"inception","family":"Mercury","kind":"llm","releaseDate":"2026-09-08","status":"ga","summary":"Diffusion language model that refines many tokens in parallel.","description":"Mercury 2.5 from Inception is a diffusion LLM: rather than writing one token at a time, it drafts and refines many tokens in parallel, which Inception says gives very high throughput.","useCases":["realtime","high-volume","coding"],"strengths":["Parallel (diffusion) generation","High throughput (vendor-reported)"],"technologies":["reasoning","diffusion-lm"],"openWeights":false,"architecture":{"type":"Diffusion LLM"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":38,"links":{"website":"https://www.inceptionlabs.ai"},"sources":[],"providerName":"Inception","blendedPrice":null},{"id":"muse-image","name":"Muse Image","providerId":"meta","family":"Muse","kind":"image","releaseDate":"2026-07-07","status":"ga","summary":"Meta’s text-to-image model from the Muse family.","description":"Muse Image is Meta’s text-to-image model from the Muse family, accepting text and reference images.","useCases":["image-generation","high-volume"],"strengths":["Muse family","Reference-image input"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToImage":1269,"arenaImageEdit":1400},"benchmarksAsOf":{"arenaTextToImage":"2026-10-26","arenaImageEdit":"2026-10-27"},"popularity":38,"links":{"website":"https://ai.meta.com"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Meta","blendedPrice":null},{"id":"sonic-3.6","name":"Sonic 3.6","providerId":"cartesia","family":"Sonic","kind":"speech","releaseDate":"2026-08-17","status":"ga","summary":"Low-latency streaming TTS for voice agents.","description":"Sonic 3.6 is Cartesia’s streaming text-to-speech model, built for real-time voice agents on Cartesia’s state-space-model research.","useCases":["realtime","agents","multimodal"],"strengths":["Built for real-time agents","Streaming output"],"technologies":["realtime-voice","hybrid-ssm"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["audio"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":38,"links":{"website":"https://cartesia.ai"},"sources":[],"providerName":"Cartesia","blendedPrice":null},{"id":"seed-2.1-turbo","name":"Seed 2.1 Turbo","providerId":"bytedance","family":"Seed 2","kind":"llm","releaseDate":"2026-08-12","status":"ga","summary":"ByteDance’s multimodal model for coding and long-horizon agents.","description":"Seed 2.1 Turbo is a ByteDance Seed model for coding and long agent workflows: delivering software end to end, carrying out multi-step tasks and understanding images and video.","useCases":["coding","agents","vision","multimodal"],"strengths":["SWE-bench Pro 57% (vendor-reported)","Video input"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"reasoningScore":43.8,"codingScore":29.7,"agentScore":28.6,"mathScore":39.5,"toolCallingScore":19.9,"visionScore":30.4,"sweBenchPro":57},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31","sweBenchPro":"2026-10-31"},"popularity":36,"links":{"website":"https://seed.bytedance.com"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard"],"providerName":"ByteDance Seed","blendedPrice":null},{"id":"qwen-image-2.1","name":"Qwen-Image-2.1","providerId":"alibaba","family":"Qwen-Image","kind":"image","releaseDate":"2026-09-20","status":"ga","summary":"Open-weight Qwen image model you can run locally.","description":"Qwen-Image-2.1 is an open-weight image generation model from Alibaba’s Qwen team that can be downloaded and run locally (e.g. in ComfyUI).","useCases":["image-generation","on-device"],"strengths":["Open weights","Run locally"],"technologies":["diffusion"],"openWeights":true,"huggingFaceId":"Qwen/Qwen-Image-2.1","architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToImage":1225,"arenaImageEdit":1362},"benchmarksAsOf":{"arenaTextToImage":"2026-10-26","arenaImageEdit":"2026-10-27"},"popularity":36,"links":{"website":"https://qwen.ai","huggingFace":"https://huggingface.co/Qwen/Qwen-Image-2.1"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Alibaba (Qwen)","blendedPrice":null},{"id":"gemini-robotics-2","name":"Gemini Robotics 2","providerId":"google","family":"Gemini Robotics","kind":"robotics","releaseDate":"2026-07-30","status":"limited","summary":"Whole-body humanoid control from vision and language.","description":"Gemini Robotics 2 is Google DeepMind’s vision-language-action model family for robots from bi-arm systems to full humanoids, adding whole-body control (walking, crouching, manipulation) and multi-robot coordination. The companion Gemini Robotics ER 2 embodied-reasoning model is in preview in Google AI Studio; the VLA and On-Device 2 models go to early-access partners.","useCases":["agents","vision","multimodal"],"strengths":["Whole-body humanoid control","ER 2 reasoning model in AI Studio","On-device variant"],"technologies":["vla","native-multimodal"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image","video"],"output":["action","text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":36,"links":{"website":"https://deepmind.google/models/gemini-robotics/"},"sources":["https://deepmind.google/models/gemini-robotics/","https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/"],"providerName":"Google","blendedPrice":null},{"id":"gpt-oss-safeguard-20b","name":"gpt-oss-safeguard-20b","providerId":"openai","family":"gpt-oss","kind":"safety","releaseDate":"2025-10-29","status":"ga","summary":"Open safety-reasoning model that applies your own written policy.","description":"gpt-oss-safeguard-20b is an open-weight OpenAI safety model built on gpt-oss-20b. Rather than using a fixed taxonomy, it reasons over a policy you write and supply at inference time, for content classification, LLM filtering and trust and safety labelling.","useCases":["enterprise","extraction","reasoning"],"strengths":["Bring-your-own policy","Explains its decisions","Apache 2.0"],"technologies":["guardrails","reasoning","moe"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"openai/gpt-oss-safeguard-20b","architecture":{"type":"MoE","totalParams":"21B","activeParams":"3.6B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":36,"links":{"website":"https://openai.com/index/introducing-gpt-oss-safeguard/","huggingFace":"https://huggingface.co/openai/gpt-oss-safeguard-20b"},"sources":[],"providerName":"OpenAI","blendedPrice":null},{"id":"granite-4.2-8b","name":"Granite 4.2 8B","providerId":"ibm","family":"Granite 4","kind":"llm","releaseDate":"2026-08-31","status":"ga","summary":"Apache-licensed small reasoning model for governed enterprise use.","description":"Granite 4.2 8B is a dense IBM reasoning model for maths, code, multilingual chat and agents, with full, low-effort and no-thinking modes.","useCases":["on-device","enterprise","extraction","high-volume"],"strengths":["Apache 2.0 with IP indemnity","Tiny footprint"],"technologies":["reasoning","transformer"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"ibm-granite/granite-4.2-8b","architecture":{"type":"Dense","totalParams":"8B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":17,"codingScore":2.9,"agentScore":2.6,"mathScore":20.8,"toolCallingScore":2.3,"sweBenchPro":19.1,"gpqa":65.7,"tauBenchAirline":58.3,"arenaText":1290},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaText":"2026-10-28"},"popularity":35,"links":{"website":"https://www.ibm.com/granite","huggingFace":"https://huggingface.co/ibm-granite/granite-4.2-8b"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"IBM","blendedPrice":null},{"id":"voyage-multimodal-3.5","name":"voyage-multimodal-3.5","providerId":"voyage","family":"Voyage Multimodal","kind":"embedding","releaseDate":"2026-07-27","status":"ga","summary":"Embeds text, images and interleaved content into one space.","description":"voyage-multimodal-3.5 vectorises text and images — including content that interleaves them, such as screenshots of documents and slides — into a single vector space for multimodal retrieval.","useCases":["vision","multimodal","research"],"strengths":["Interleaved text + image inputs","Document screenshots without OCR"],"technologies":["vector-search","rag","native-multimodal"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":34,"links":{"website":"https://www.voyageai.com","docs":"https://docs.voyageai.com/docs/multimodal-embeddings"},"sources":[],"providerName":"Voyage AI (MongoDB)","blendedPrice":null},{"id":"embed-v5.0-pro","name":"Embed 5 Pro","providerId":"cohere","family":"Embed 5","kind":"embedding","releaseDate":"2026-09-30","status":"ga","summary":"Cohere’s highest-quality multimodal embedding model, for indexing.","description":"Embed 5 Pro is the quality tier of Cohere’s Embed 5, meant for indexing documents offline. Both Embed 5 tiers share one embedding space, so documents indexed with Pro can be queried with Fast. They embed text, images and mixed pages in 100+ languages with a 128K-token input, and return Matryoshka vectors from 256 to 2048 dimensions. It is on the Cohere API and cloud marketplaces, and can be self-hosted.","useCases":["multimodal","enterprise","long-context"],"strengths":["Text, image and mixed inputs","128K-token input","Shares a vector space with Embed 5 Fast"],"technologies":["vector-search","rag"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null,"note":"$0.12 per 1M text tokens; $0.40 per 1M image tokens"},"benchmarks":{},"benchmarksAsOf":{},"popularity":34,"links":{"website":"https://cohere.com/embed"},"sources":["https://cohere.com/blog/embed-5","https://docs.cohere.com/changelog/embed-v5","https://docs.cohere.com/docs/cohere-embed"],"providerName":"Cohere","blendedPrice":null},{"id":"krea-3","name":"Krea 3","providerId":"krea","family":"Krea 3","kind":"image","releaseDate":"2026-10-15","status":"ga","summary":"Krea’s fast image-editing model, native 2K output.","description":"Krea 3 edits images from a text instruction and up to six reference images, returning native 2K (or 1K) results in seconds. Most edits finish in under 8 seconds. It is the default image model in the Krea app and is available in the Krea API; an Ultra tier is offered on Max plans and above.","useCases":["image-generation"],"strengths":["Fast edits","Native 2K output","Multiple reference images"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null,"note":"$0.06 per image at 2K; $0.03 at 1K"},"benchmarks":{},"benchmarksAsOf":{},"popularity":34,"links":{"website":"https://www.krea.ai"},"sources":["https://www.krea.ai/blog/introducing-krea-3","https://www.krea.ai/docs/user-guide/features/krea-3"],"providerName":"Krea","blendedPrice":null},{"id":"ltx-2.5","name":"LTX-2.5","providerId":"lightricks","family":"LTX","kind":"video","releaseDate":"2026-08-11","status":"ga","summary":"Open-weight video model you can run locally.","description":"LTX-2.5 is Lightricks’ open-weight video generation model (Fast and Pro tiers), runnable locally in ComfyUI or via API.","useCases":["multimodal","on-device"],"strengths":["Open weights","Fast tier for iteration"],"technologies":["diffusion"],"openWeights":true,"huggingFaceId":"Lightricks/LTX-2.5","architecture":{},"modalities":{"input":["text","image"],"output":["video"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":34,"links":{"website":"https://ltx.studio","huggingFace":"https://huggingface.co/Lightricks/LTX-2.5"},"sources":[],"providerName":"Lightricks","blendedPrice":null},{"id":"voxtral-small","name":"Voxtral Small","providerId":"mistral","family":"Voxtral","kind":"transcription","releaseDate":"2025-07-15","status":"ga","summary":"Open-weight Mistral model that transcribes and understands audio.","description":"Voxtral Small extends Mistral Small 3 with audio input, handling transcription, translation and questions about audio while keeping its text abilities. Weights are Apache 2.0.","useCases":["extraction","multimodal","on-device"],"strengths":["Open weights","Audio Q&A, not just transcripts"],"technologies":[],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"mistralai/Voxtral-Small-24B-2507","architecture":{"totalParams":"24B"},"modalities":{"input":["text","audio","file"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":34,"links":{"website":"https://mistral.ai/news/voxtral","huggingFace":"https://huggingface.co/mistralai/Voxtral-Small-24B-2507"},"sources":[],"providerName":"Mistral AI","blendedPrice":null},{"id":"v-jepa-2","name":"V-JEPA 2","providerId":"meta","family":"V-JEPA","kind":"world-model","releaseDate":"2025-06-11","status":"ga","summary":"Meta’s open JEPA world model for video understanding, prediction and robot planning.","description":"V-JEPA 2 is a self-supervised video world model from Meta FAIR built on the Joint-Embedding Predictive Architecture: instead of generating pixels it predicts representations of masked and future video. Pre-trained on over a million hours of internet video, it is used as a video encoder for understanding and retrieval; the action-conditioned V-JEPA 2-AC variant, fine-tuned on about 62 hours of robot video, plans zero-shot pick-and-place on real robot arms.","useCases":["vision","agents","research"],"strengths":["Open weights (Apache 2.0)","Predicts in latent space, not pixels","Zero-shot robot planning (V-JEPA 2-AC)"],"technologies":["jepa"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"facebook/vjepa2-vitg-fpc64-256","architecture":{"type":"JEPA (ViT-g encoder)","totalParams":"1B"},"modalities":{"input":["video","image"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":34,"links":{"website":"https://ai.meta.com/vjepa/","docs":"https://github.com/facebookresearch/vjepa2","huggingFace":"https://huggingface.co/facebook/vjepa2-vitg-fpc64-256"},"sources":["https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/","https://huggingface.co/facebook/vjepa2-vitg-fpc64-256"],"providerName":"Meta","blendedPrice":null},{"id":"nemotron-3.5-bolt","name":"Nemotron 3.5 Bolt","providerId":"nvidia","family":"Nemotron 3.5","kind":"llm","releaseDate":"2026-10-15","status":"ga","summary":"Compact open vision-language MoE for running agents locally.","description":"Nemotron 3.5 Bolt is NVIDIA’s smaller open vision-language model, a hybrid Mamba-Transformer mixture of experts (74B total, 9B active) compressed from Nemotron 3.5 Super VL. NVIDIA publishes a full-precision checkpoint and a 24GB build sized for a single workstation GPU or DGX Spark. It reads text, images and video, and thinking can be switched on or off.","useCases":["agents","vision","on-device","long-context"],"strengths":["Fits in 24GB (quantized build)","Image and video input","Open weights"],"technologies":["hybrid-ssm","reasoning","moe"],"openWeights":true,"license":"OpenMDW 1.1","huggingFaceId":"nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB","architecture":{"type":"MoE","totalParams":"74B","activeParams":"9B"},"contextWindow":262144,"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"reasoningScore":28.1,"codingScore":15.2,"agentScore":11.6,"mathScore":23.4,"toolCallingScore":5.1,"visionScore":10.9},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31"},"popularity":32,"links":{"huggingFace":"https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB"},"sources":["https://research.nvidia.com/labs/nemotron/Nemotron-3.5-Bolt-24GB/","https://llm-stats.com/leaderboards/llm-leaderboard"],"providerName":"NVIDIA","blendedPrice":null},{"id":"pi-0.5","name":"π0.5","providerId":"physical-intelligence","family":"π0","kind":"robotics","releaseDate":"2025-04-22","status":"ga","summary":"Open VLA from Physical Intelligence with strong open-world generalisation.","description":"π0.5 upgrades Physical Intelligence’s π0 flow-matching VLA with better open-world generalisation (trained with knowledge insulation). Base and DROID/LIBERO checkpoints are open-sourced in the openpi repo for inference and fine-tuning on your own robot.","useCases":["agents","vision","on-device"],"strengths":["Open weights (Apache 2.0)","Fine-tune with LoRA on one GPU","PyTorch + JAX"],"technologies":["vla","diffusion","fine-tuning"],"openWeights":true,"license":"Apache 2.0","architecture":{},"modalities":{"input":["text","image"],"output":["action"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":32,"links":{"website":"https://www.physicalintelligence.company","docs":"https://github.com/Physical-Intelligence/openpi"},"sources":["https://github.com/Physical-Intelligence/openpi"],"providerName":"Physical Intelligence","blendedPrice":null},{"id":"qwen3.8-omni-flash","name":"Qwen3.8 Omni Flash","providerId":"alibaba","family":"Qwen3.8","kind":"llm","releaseDate":"2026-09-21","status":"ga","summary":"Omni-modal reasoning model with native audio and video understanding.","description":"Qwen3.8 Omni Flash is Alibaba’s omni-modal reasoning model and, per Alibaba, the first Qwen model designed around agent use with built-in audio and video understanding. It suits analysing and summarising audio and video.","useCases":["multimodal","agents","high-volume","long-context"],"strengths":["Native audio + video input","Agent-oriented design"],"technologies":["native-multimodal","reasoning"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image","audio","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"reasoningScore":31.6,"agentScore":22.2},"benchmarksAsOf":{"reasoningScore":"2026-10-31","agentScore":"2026-10-31"},"popularity":30,"links":{"website":"https://qwen.ai"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard"],"providerName":"Alibaba (Qwen)","blendedPrice":null},{"id":"step-5-preview","name":"Step 5 Preview","providerId":"stepfun","family":"Step 5","kind":"llm","releaseDate":"2026-10-08","status":"preview","summary":"StepFun’s new agentic flagship — 600B MoE with video input.","description":"Step 5 Preview is StepFun’s new flagship for agent work, a sparse mixture of experts with 27B active out of 600B parameters. StepFun highlights software engineering and professional tasks.","useCases":["agents","coding","multimodal"],"strengths":["Video input","Agent-focused"],"technologies":["reasoning","moe"],"openWeights":false,"architecture":{"type":"MoE","totalParams":"600B","activeParams":"27B"},"modalities":{"input":["text","image","video"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":3.75,"usageRank":7,"asOf":"2026-10-31","topTasks":[{"label":"Summarization","share":13.2},{"label":"Q&A & Knowledge","share":4.5}],"tasksAsOf":"2026-10-30"},"benchmarks":{"arenaText":1448,"arenaWebdev":1571},"benchmarksAsOf":{"arenaText":"2026-10-28","arenaWebdev":"2026-10-28"},"popularity":30,"links":{},"sources":["https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"StepFun","blendedPrice":null},{"id":"granite-4.0-h-micro","name":"Granite 4.0 H Micro","providerId":"ibm","family":"Granite 4","kind":"llm","releaseDate":"2025-10-02","status":"ga","summary":"Tiny hybrid Mamba-2/Transformer model for long-input enterprise tasks.","description":"Granite 4.0 H Micro is a 3B-parameter IBM model. The “H” models mix Mamba-2 state-space layers with attention, which cuts memory use on long inputs and many concurrent sessions. Apache 2.0.","useCases":["on-device","extraction","enterprise","high-volume"],"strengths":["Hybrid Mamba-2 architecture","Apache 2.0","Runs on modest GPUs"],"technologies":["hybrid-ssm"],"openWeights":true,"license":"Apache 2.0","huggingFaceId":"ibm-granite/granite-4.0-h-micro","architecture":{"type":"Hybrid Mamba-2/Transformer (dense)","totalParams":"3B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":30,"links":{"website":"https://www.ibm.com/granite","huggingFace":"https://huggingface.co/ibm-granite/granite-4.0-h-micro"},"sources":[],"providerName":"IBM","blendedPrice":null},{"id":"solar-pro-4","name":"Solar Pro 4","providerId":"upstage","family":"Solar","kind":"llm","releaseDate":"2026-08-10","status":"ga","summary":"Upstage’s reasoning model for long, document-heavy and agentic work.","description":"Solar Pro 4 is Upstage’s larger language model, built for long agent tasks, office productivity and document-heavy work.","useCases":["agents","long-context","extraction","enterprise"],"strengths":["Document-heavy tasks","Agentic workflows"],"technologies":["reasoning"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.32,"usageRank":38,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":33.5,"codingScore":20.4,"agentScore":10.8,"mathScore":32.1,"toolCallingScore":8.9,"arenaText":1388,"arenaWebdev":1370},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","arenaText":"2026-10-28","arenaWebdev":"2026-10-28"},"popularity":30,"links":{"website":"https://www.upstage.ai"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Upstage","blendedPrice":null},{"id":"gpt-6-luna-decisions","name":"GPT-6 Luna Decisions","providerId":"openai","family":"GPT-6","kind":"decision","releaseDate":"2026-10-06","status":"ga","summary":"GPT-6 Luna served through OpenAI’s Decisions API for typed, probabilistic answers.","description":"GPT-6 Luna Decisions is GPT-6 Luna offered through OpenAI’s Decisions API, in public beta from October 6, 2026 and generally available since October 22, 2026 (up to 400 questions per request). Instead of writing text, it reads text or images and returns typed answers: the probability that a statement is true, a pick from fixed options with a probability for each, or a score on a scale.","useCases":["extraction","agents","long-context","enterprise"],"strengths":["Image input","Typed decisions"],"technologies":["system-one"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["decision"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":30,"links":{"website":"https://openai.com","docs":"https://developers.openai.com/api/docs/guides/decisions"},"sources":["https://developers.openai.com/api/docs/changelog"],"providerName":"OpenAI","blendedPrice":null},{"id":"nemotron-3-embed-1b","name":"Nemotron 3 Embed 1B","providerId":"nvidia","family":"Nemotron Embed","kind":"embedding","releaseDate":"2026-07-16","status":"ga","summary":"Small open embedding model for high-throughput enterprise retrieval.","description":"NVIDIA Nemotron 3 Embed 1B is an open text embedding model optimised for high-throughput, low-latency retrieval — enterprise search, RAG, code retrieval and agentic retrieval.","useCases":["high-volume","enterprise","on-device"],"strengths":["1B parameters — cheap to self-host","Open weights"],"technologies":["vector-search","rag"],"openWeights":true,"architecture":{"totalParams":"1B"},"modalities":{"input":["text"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":30,"links":{"website":"https://build.nvidia.com"},"sources":[],"providerName":"NVIDIA","blendedPrice":null},{"id":"embeddinggemma-2","name":"EmbeddingGemma 2","providerId":"google","family":"Gemma 4","kind":"embedding","releaseDate":"2026-10-06","status":"ga","summary":"Open 740M multimodal embedding model that runs on phones.","description":"EmbeddingGemma 2 is Google’s open embedding model built on Gemma 4. It maps text, code, images, audio and video into one embedding space with an 8K-token input (4× EmbeddingGemma 1). It is modular — text-only use needs about 270M parameters, with optional 170M vision and 300M audio encoders — and supports Matryoshka truncation from 768 to 512, 256 or 128 dimensions. Google reports best-in-class results among sub-1B multimodal embedders on MTEB Code and MAEB (vendor-reported).","useCases":["on-device","multimodal","coding"],"strengths":["Open weights (Apache-2.0)","Text, image, audio and video","Runs on-device"],"technologies":["vector-search","rag","native-multimodal"],"openWeights":true,"license":"Apache-2.0","huggingFaceId":"google/embeddinggemma-2","architecture":{"totalParams":"740M"},"modalities":{"input":["text","image","audio","video"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":30,"links":{"website":"https://ai.google.dev/gemma","huggingFace":"https://huggingface.co/google/embeddinggemma-2"},"sources":["https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/"],"providerName":"Google","blendedPrice":null},{"id":"krea-2-large","name":"Krea 2 Large","providerId":"krea","family":"Krea 2","kind":"image","releaseDate":"2026-05-26","status":"ga","summary":"Krea’s largest image model, with a rawer, more textured look.","description":"Krea 2 Large is Krea’s most capable image model, over twice the size of Krea 2 Medium. Krea says lighter post-training gives its images a rawer, more textured look.","useCases":["image-generation"],"strengths":["Textured, less “AI-polished” look","Text and image input"],"technologies":["diffusion"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaTextToImage":1115},"benchmarksAsOf":{"arenaTextToImage":"2026-10-26"},"popularity":30,"links":{"website":"https://www.krea.ai"},"sources":["https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Krea","blendedPrice":null},{"id":"happyhorse-1.1","name":"HappyHorse 1.1","providerId":"alibaba","family":"HappyHorse","kind":"video","releaseDate":"2026-06-22","status":"ga","summary":"Alibaba video model that works from text, a start image or reference images.","description":"HappyHorse 1.1 is an Alibaba video generation model that makes short clips from a text prompt, one starting image or several reference images.","useCases":["multimodal"],"strengths":["Reference-image conditioning","Text- or image-to-video"],"technologies":["diffusion"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["video"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":30,"links":{},"sources":[],"providerName":"Alibaba (Qwen)","blendedPrice":null},{"id":"qwen-audio-3.1-tts-plus","name":"Qwen-Audio-3.1-TTS-Plus","providerId":"alibaba","family":"Qwen Audio","kind":"speech","releaseDate":"2026-09-24","status":"ga","summary":"Alibaba’s Qwen text-to-speech model.","description":"Qwen-Audio-3.1-TTS-Plus is Alibaba’s text-to-speech model in the Qwen family.","useCases":["multimodal","high-volume"],"strengths":["Qwen audio family"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["audio"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":30,"links":{"website":"https://qwen.ai"},"sources":[],"providerName":"Alibaba (Qwen)","blendedPrice":null},{"id":"gemini-3.5-transcribe","name":"Gemini 3.5 Transcribe","providerId":"google","family":"Gemini Transcribe","kind":"transcription","releaseDate":"2026-08-26","status":"ga","summary":"Google’s dedicated transcription model.","description":"Gemini 3.5 Transcribe is Google’s dedicated speech-to-text model in the Gemini family.","useCases":["extraction","multimodal"],"strengths":["Gemini family"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["audio"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":30,"links":{"website":"https://ai.google.dev"},"sources":[],"providerName":"Google","blendedPrice":null},{"id":"lyria-3-pro","name":"Lyria 3 Pro","providerId":"google","family":"Lyria","kind":"music","releaseDate":"2026-03-30","status":"preview","summary":"Google’s full-length song generation model.","description":"Lyria 3 is Google’s family of music generation models in the Gemini API; Lyria 3 Pro generates full-length, high-quality 48 kHz songs from text and image prompts.","useCases":["multimodal","writing"],"strengths":["Full-length songs","48 kHz audio"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["audio","text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":30,"links":{"website":"https://deepmind.google/models/lyria/"},"sources":[],"providerName":"Google","blendedPrice":null},{"id":"gr00t-n1.7","name":"Isaac GR00T N1.7","providerId":"nvidia","family":"Isaac GR00T","kind":"robotics","releaseDate":"2026-04-17","status":"preview","summary":"Open 3B cross-embodiment VLA for humanoids and manipulators.","description":"GR00T N1.7 (early access) is NVIDIA’s open vision-language-action model. A Cosmos-Reason2 VLM backbone reads camera frames and a language instruction; a flow-matching diffusion-transformer action head outputs continuous motor commands for the robot’s embodiment. Post-train it on real or synthetic data for a specific robot.","useCases":["agents","vision","on-device"],"strengths":["Open weights, commercial use allowed","Cross-embodiment","3B parameters"],"technologies":["vla","diffusion"],"openWeights":true,"license":"NVIDIA Open Model License","huggingFaceId":"nvidia/GR00T-N1.7-3B","architecture":{"type":"VLM + flow-matching DiT action head","totalParams":"3B"},"modalities":{"input":["text","image"],"output":["action"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":30,"links":{"website":"https://developer.nvidia.com/isaac/gr00t","huggingFace":"https://huggingface.co/nvidia/GR00T-N1.7-3B"},"sources":["https://huggingface.co/nvidia/GR00T-N1.7-3B"],"providerName":"NVIDIA","blendedPrice":null},{"id":"hy4-preview","name":"Hy4 Preview","providerId":"tencent","family":"Hy","kind":"llm","releaseDate":"2026-08-28","status":"preview","summary":"Tencent’s open-weight 770B MoE (49B active) for coding agents and tool use.","description":"Hy4 Preview is a Tencent mixture of experts with 49B of 770B parameters active, built for coding agents, complex tool use and productivity work. Weights are Apache-2.0 on Hugging Face.","useCases":["coding","agents","long-context"],"strengths":["Open weights (Apache-2.0)","Coding agents"],"technologies":["reasoning","moe"],"openWeights":true,"license":"Apache-2.0","huggingFaceId":"tencent/Hy4-preview","architecture":{"type":"MoE","totalParams":"770B","activeParams":"49B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":4.62,"usageRank":3,"asOf":"2026-10-31","topTasks":[{"label":"Research & Reports","share":8.9},{"label":"File I/O","share":8.7},{"label":"Frontend & UI","share":6.4}],"tasksAsOf":"2026-10-30"},"benchmarks":{"reasoningScore":49.1,"codingScore":37.4,"agentScore":35,"mathScore":38.6,"toolCallingScore":29.1,"sweBenchPro":65.7,"gpqa":90,"tauBenchAirline":75.8,"arenaWebdev":1631},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","sweBenchPro":"2026-10-31","gpqa":"2026-10-30","tauBenchAirline":"2026-10-30","arenaWebdev":"2026-10-28"},"popularity":28,"links":{"website":"https://hunyuan.tencent.com","huggingFace":"https://huggingface.co/tencent/Hy4-preview"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Tencent","blendedPrice":null},{"id":"embed-v5.0-fast","name":"Embed 5 Fast","providerId":"cohere","family":"Embed 5","kind":"embedding","releaseDate":"2026-09-30","status":"ga","summary":"Lower-latency, cheaper Embed 5 tier for live queries.","description":"Embed 5 Fast is the low-latency tier of Cohere’s Embed 5, meant for embedding live queries at lower cost. Both Embed 5 tiers share one embedding space, so documents indexed with Pro can be queried with Fast. They embed text, images and mixed pages in 100+ languages with a 128K-token input, and return Matryoshka vectors from 256 to 2048 dimensions. Cohere reports slightly lower quality than Pro (vendor-reported).","useCases":["multimodal","enterprise","long-context"],"strengths":["Lower price and latency","Same vector space as Embed 5 Pro","128K-token input"],"technologies":["vector-search","rag"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null,"note":"$0.08 per 1M text tokens; $0.40 per 1M image tokens"},"benchmarks":{},"benchmarksAsOf":{},"popularity":28,"links":{"website":"https://cohere.com/embed"},"sources":["https://cohere.com/blog/embed-5","https://docs.cohere.com/changelog/embed-v5","https://docs.cohere.com/docs/cohere-embed"],"providerName":"Cohere","blendedPrice":null},{"id":"mai-image-2.7-preview","name":"MAI-Image-2.7 Preview","providerId":"microsoft","family":"MAI-Image","kind":"image","releaseDate":"2026-10-27","status":"preview","summary":"Preview of Microsoft AI’s next image model, editing only for now.","description":"MAI-Image-2.7 is the next version of Microsoft AI’s in-house image model, launched as a preview that edits existing images from text instructions; Microsoft says text-to-image generation will follow. It can be tried in MAI Playground; no API or price has been announced.","useCases":["image-generation"],"strengths":["Image editing","Improves on MAI-Image-2.6 (vendor-reported)"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"arenaImageEdit":1471},"benchmarksAsOf":{"arenaImageEdit":"2026-10-27"},"popularity":28,"links":{"website":"https://microsoft.ai"},"sources":["https://microsoft.ai/news/mai-image-2-7-debut/","https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset"],"providerName":"Microsoft AI","blendedPrice":null},{"id":"recraft-v4.1-utility","name":"Recraft V4.1 Utility","providerId":"recraft","family":"Recraft V4","kind":"image","releaseDate":"2026-05-14","status":"ga","summary":"Recraft’s general-purpose image model, strong for design work.","description":"Recraft V4.1 Utility is Recraft’s general-purpose image model for design and marketing images in several aspect ratios. Recraft also offers Pro and vector variants.","useCases":["image-generation"],"strengths":["Design-oriented output","Pro and vector variants"],"technologies":["diffusion"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":28,"links":{"website":"https://www.recraft.ai"},"sources":[],"providerName":"Recraft","blendedPrice":null},{"id":"longcat-2.0","name":"LongCat 2.0","providerId":"meituan","family":"LongCat","kind":"llm","releaseDate":"2026-07-20","status":"ga","summary":"Open 1.6T-parameter MoE for repository-scale coding.","description":"LongCat 2.0 is Meituan’s sparse mixture of experts, with 48B of 1.6T parameters active, aimed at coding, repository-wide changes, long problem solving and agents.","useCases":["coding","agents","long-context"],"strengths":["Open weights","Repository-scale coding"],"technologies":["reasoning","moe"],"openWeights":true,"huggingFaceId":"meituan-longcat/LongCat-2.0","architecture":{"type":"MoE","totalParams":"1.6T","activeParams":"48B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":26,"links":{"website":"https://longcat.ai","huggingFace":"https://huggingface.co/meituan-longcat/LongCat-2.0"},"sources":[],"providerName":"Meituan (LongCat)","blendedPrice":null},{"id":"pplx-embed-v1-4b","name":"pplx-embed-v1 4B","providerId":"perplexity","family":"pplx-embed","kind":"embedding","releaseDate":"2026-03-15","status":"ga","summary":"Perplexity’s web-scale retrieval embedding.","description":"pplx-embed-v1-4B is Perplexity’s larger text embedding model for real-world, web-scale dense retrieval; a 0.6B sibling trades quality for cost.","useCases":["research","high-volume"],"strengths":["Built for web-scale search","0.6B sibling for lower cost"],"technologies":["vector-search","rag"],"openWeights":false,"architecture":{"totalParams":"4B"},"modalities":{"input":["text"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":26,"links":{"website":"https://www.perplexity.ai"},"sources":[],"providerName":"Perplexity","blendedPrice":null},{"id":"mai-transcribe-2","name":"MAI-Transcribe-2","providerId":"microsoft","family":"MAI-Transcribe","kind":"transcription","releaseDate":"2026-09-03","status":"ga","summary":"Microsoft AI’s in-house speech-to-text model.","description":"MAI-Transcribe-2 is Microsoft AI’s in-house speech-to-text model for batch transcription.","useCases":["extraction","high-volume","enterprise"],"strengths":["In-house Microsoft model","Batch transcription"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["audio"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":26,"links":{"website":"https://microsoft.ai"},"sources":[],"providerName":"Microsoft AI","blendedPrice":null},{"id":"ling-3.1-flash","name":"Ling 3.1 Flash","providerId":"inclusionai","family":"Ling","kind":"llm","releaseDate":"2026-10-02","status":"ga","summary":"Hybrid-reasoning MoE from inclusionAI (25B active / 560B).","description":"Ling 3.1 Flash is a hybrid reasoning mixture of experts from inclusionAI, Ant Group’s AI lab, with 25B of 560B parameters active.","useCases":["reasoning","agents","realtime","long-context"],"strengths":["Hybrid reasoning","25B active parameters"],"technologies":["reasoning","moe"],"openWeights":false,"architecture":{"type":"MoE","totalParams":"560B","activeParams":"25B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.13,"usageRank":52,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"gpqa":81.3,"tauBenchAirline":79.3},"benchmarksAsOf":{"gpqa":"2026-10-30","tauBenchAirline":"2026-10-30"},"popularity":24,"links":{"website":"https://github.com/inclusionAI"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings"],"providerName":"inclusionAI (Ant Group)","blendedPrice":null},{"id":"microsoft-decision-1","name":"Microsoft-Decision-1","providerId":"microsoft","family":"Microsoft Decision","kind":"decision","releaseDate":"2026-10-09","status":"ga","summary":"Microsoft’s small decision-scoring model: calibrated probabilities over fixed options.","description":"Microsoft-Decision-1 is a decision model post-trained from Alibaba’s open-weight Qwen3.5-9B. Given text and a fixed set of answer options — yes/no, multiple choice, a rating, or a rubric for grading AI responses and agent actions — it returns a calibrated probability for each option in a single pass. Microsoft reports the highest accuracy in its own 36-benchmark comparison and P50 latency about 35× faster than GPT-6 Sol (vendor-reported). Microsoft advises against using it as the sole basis for consequential decisions about people.","useCases":["extraction","realtime","high-volume","agents","enterprise"],"strengths":["Calibrated per-option probabilities","Rubric grading of agent actions"],"technologies":["system-one","fine-tuning"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["decision"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":24,"links":{"website":"https://ai.azure.com/catalog/models/Microsoft-Decision-1","docs":"https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-microsoft-decision"},"sources":["https://commandline.microsoft.com/microsoft-decision-1-model-foundry/"],"providerName":"Microsoft AI","blendedPrice":null},{"id":"inworld-realtime-tts-2","name":"Inworld Realtime TTS-2","providerId":"inworld","family":"Inworld TTS","kind":"speech","releaseDate":"2026-08-31","status":"ga","summary":"Real-time TTS for games and interactive characters.","description":"Realtime TTS-2 is Inworld’s streaming text-to-speech model for interactive applications, with a cheaper Flash variant.","useCases":["realtime","multimodal"],"strengths":["Real-time streaming","Cheaper Flash variant"],"technologies":["realtime-voice"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["audio"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":24,"links":{"website":"https://inworld.ai"},"sources":[],"providerName":"Inworld","blendedPrice":null},{"id":"mai-voice-2.1","name":"MAI-Voice-2.1","providerId":"microsoft","family":"MAI-Voice","kind":"speech","releaseDate":"2026-10-01","status":"ga","summary":"Microsoft AI’s most expressive multilingual text-to-speech model.","description":"MAI-Voice-2.1 is Microsoft AI’s highest-fidelity text-to-speech model, speaking 23 languages across 26 locales. One voice can speak every supported language with a native accent, and voice cloning from a few seconds of audio includes built-in consent checks. Microsoft offers it in Microsoft Foundry and MAI Playground.","useCases":["multimodal","writing","enterprise"],"strengths":["23 languages from one voice","Voice cloning with consent checks"],"technologies":["realtime-voice"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["audio"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":24,"links":{"website":"https://microsoft.ai"},"sources":["https://microsoft.ai/news/our-first-streaming-transcription-model/"],"providerName":"Microsoft AI","blendedPrice":null},{"id":"nemotron-3.5-content-safety","name":"Nemotron 3.5 Content Safety","providerId":"nvidia","family":"Nemotron Safety","kind":"safety","releaseDate":"2026-06-04","status":"ga","summary":"Compact 4B multimodal guardrail model for LLM and VLM traffic.","description":"Nemotron 3.5 Content Safety is a 4B-parameter NVIDIA guardrail model fine-tuned from Gemma 3 4B. It moderates both what goes into and what comes out of text and vision-language models.","useCases":["enterprise","high-volume","vision"],"strengths":["4B — cheap to run inline","Text + image","Open weights"],"technologies":["guardrails","transformer"],"openWeights":true,"huggingFaceId":"nvidia/Nemotron-3.5-Content-Safety","architecture":{"type":"Dense","totalParams":"4B"},"modalities":{"input":["text","image"],"output":["text"]},"features":{"reasoning":true,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":24,"links":{"website":"https://build.nvidia.com","huggingFace":"https://huggingface.co/nvidia/Nemotron-3.5-Content-Safety"},"sources":[],"providerName":"NVIDIA","blendedPrice":null},{"id":"nemotron-labs-dory-1.0-4b","name":"Nemotron Labs Dory 1.0 4B","providerId":"nvidia","family":"Nemotron Labs","kind":"llm","releaseDate":"2026-10-16","status":"ga","summary":"Small open looped model that reuses a shared block to think harder.","description":"Dory 1.0 4B is an open 4B-parameter model from NVIDIA’s Nemotron Labs that runs a shared middle block of layers several times per token (one to four loops; NVIDIA recommends two), so one set of weights trades speed for accuracy. It is trained on English and 43 programming languages, always reasons, and supports tool calling.","useCases":["reasoning","coding","on-device"],"strengths":["Looped (recurrent-depth) design","256K-token context","Open weights"],"technologies":["looped-lm","reasoning"],"openWeights":true,"license":"OpenMDW 1.1","huggingFaceId":"nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16","architecture":{"type":"Looped transformer","totalParams":"4B"},"contextWindow":262144,"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"reasoningScore":19.2,"codingScore":3.5,"agentScore":-2.3,"mathScore":12.2,"toolCallingScore":-3.1,"visionScore":-2.5},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31","visionScore":"2026-10-31"},"popularity":22,"links":{"huggingFace":"https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard"],"providerName":"NVIDIA","blendedPrice":null},{"id":"lfm-2.5-2.6b","name":"LFM2.5-2.6B","providerId":"liquid","family":"LFM 2.5","kind":"llm","releaseDate":"2026-08-11","status":"ga","summary":"Compact on-device reasoning model on Liquid’s hybrid architecture.","description":"LFM2.5-2.6B is a small Liquid AI reasoning model for agents, data extraction, RAG and long documents on edge devices. Liquid recommends against using it for agentic coding.","useCases":["on-device","extraction","agents","high-volume"],"strengths":["2.6B parameters","Open weights","Edge deployment"],"technologies":["reasoning"],"openWeights":true,"huggingFaceId":"LiquidAI/LFM2.5-2.6B","architecture":{"totalParams":"2.69B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{"reasoningScore":13.9,"codingScore":16.2,"agentScore":6.6,"mathScore":2.7,"toolCallingScore":0.1},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31"},"popularity":22,"links":{"website":"https://www.liquid.ai","huggingFace":"https://huggingface.co/LiquidAI/LFM2.5-2.6B"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard"],"providerName":"Liquid AI","blendedPrice":null},{"id":"jev","name":"Jev","providerId":"typesafe","family":"System One","kind":"decision","releaseDate":"2026-10-09","status":"ga","summary":"The first System One model: unstructured state in, typed probabilistic decisions out — in milliseconds.","description":"Jev is TypeSafe AI’s first System One model, made for the quick, structured decisions inside software rather than for chat. You send a state plus typed questions — Choice (pick from a list), Score (rate against a rubric) or Noul (a true/false-style judgment) — and get schema-defined answers with calibrated probabilities and a confidence value. Questions run in parallel, so adding more barely changes latency. It uses a parallel sampler and is trained with Reinforcement Learning for Calibrated Decisions (RLCD). Because it never writes free text, its outputs cannot fall outside the schema. TypeSafe reports 70–500 ms end-to-end latency and large speed and cost gains over frontier LLMs on its own workflow evals (vendor-reported). The current version, Jev 1.14 (October 14, 2026), also accepts images.","useCases":["extraction","realtime","high-volume","agents","enterprise"],"strengths":["Typed outputs that cannot mismatch the schema","Calibrated probabilities on every decision","Millisecond-scale latency (vendor-reported)"],"technologies":["system-one","constrained-decoding"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["decision"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":2.25,"usageRank":13,"asOf":"2026-10-31","topTasks":[{"label":"Classification","share":31.6},{"label":"Finance & Trading","share":15.7},{"label":"Q&A & Knowledge","share":14.1}],"tasksAsOf":"2026-10-30"},"benchmarks":{},"benchmarksAsOf":{},"popularity":22,"links":{"website":"https://typesafe.ai","docs":"https://docs.typesafe.ai/introduction"},"sources":["https://typesafe.ai/blog/introducing-system-one-models-and-jev","https://docs.typesafe.ai/introduction","https://openrouter.ai/rankings"],"providerName":"TypeSafe AI","blendedPrice":null},{"id":"clef","name":"Clef","providerId":"cloudflare","family":"Clef","kind":"decision","releaseDate":"2026-10-01","status":"ga","summary":"Cloudflare’s open-source 27B multimodal decision model for high-precision decisions.","description":"Clef is a 27B decision model from Cloudflare’s Workers AI team, fine-tuned from Qwen3.8-27B. It reads a state (text or JSON, plus up to four images) and typed questions, and returns a probability for each allowed answer. Weights are Apache-2.0 on Hugging Face; it is also served on Workers AI (@cf/cloudflare/clef). Cloudflare reports leading 7 of 10 decision benchmarks within the Clef family and a 209 ms median latency (vendor-reported).","useCases":["extraction","agents","vision","enterprise"],"strengths":["Open weights (Apache-2.0)","Image input","Served on Workers AI"],"technologies":["system-one","fine-tuning"],"openWeights":true,"license":"Apache-2.0","huggingFaceId":"Cloudflare/clef","architecture":{"totalParams":"27B"},"modalities":{"input":["text","image"],"output":["decision"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":22,"links":{"website":"https://developers.cloudflare.com/workers-ai/","huggingFace":"https://huggingface.co/Cloudflare/clef"},"sources":["https://developers.cloudflare.com/changelog/post/2026-10-01-clef-workers-ai/"],"providerName":"Cloudflare","blendedPrice":null},{"id":"hy-image-3.5-preview","name":"Hy Image 3.5 Preview","providerId":"tencent","family":"Hy Image","kind":"image","releaseDate":"2026-10-05","status":"preview","summary":"Tencent’s unified image generation and multi-turn editing model.","description":"Hy Image 3.5 Preview is Tencent’s combined image generator and editor. A single endpoint covers text-to-image, image-to-image and multi-turn edits, and it can take up to 20 reference images.","useCases":["image-generation","multimodal"],"strengths":["Up to 20 reference images","Generation and editing in one model"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["image"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":22,"links":{"website":"https://hunyuan.tencent.com"},"sources":[],"providerName":"Tencent","blendedPrice":null},{"id":"nova-2.5-sonic","name":"Nova 2.5 Sonic","providerId":"amazon","family":"Nova","kind":"speech","releaseDate":"2026-10-05","status":"ga","summary":"Amazon’s speech-to-speech model for real-time voice agents on Bedrock.","description":"Nova 2.5 Sonic is Amazon’s speech-to-speech model for natural, real-time voice agents. AWS says it improves reasoning, instruction following and tool-calling accuracy and lowers latency compared with Nova 2 Sonic, without publishing figures. It offers expressive voices in seven languages, controllable turn-taking, mixed voice and text in one session, asynchronous tool calling and a 256K-token context, on Amazon Bedrock in four regions.","useCases":["realtime","agents","chat"],"strengths":["Speech in, speech out","Asynchronous tool calling","256K-token context"],"technologies":["realtime-voice"],"openWeights":false,"architecture":{},"modalities":{"input":["audio","text"],"output":["audio","text"]},"features":{"reasoning":false,"toolCalling":true,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":22,"links":{"website":"https://aws.amazon.com/bedrock/nova/"},"sources":["https://aws.amazon.com/about-aws/whats-new/2026/10/amazon-nova-2.5-sonic/"],"providerName":"Amazon","blendedPrice":null},{"id":"solar-mini-4","name":"Solar Mini 4","providerId":"upstage","family":"Solar","kind":"llm","releaseDate":"2026-09-23","status":"ga","summary":"Upstage’s compact 35B MoE (3B active) for fast agentic work.","description":"Solar Mini 4 is Upstage’s small, efficient model: a 35B mixture of experts with 3B active parameters, designed for agent use cases where quick responses matter.","useCases":["agents","high-volume","realtime","long-context"],"strengths":["3B active parameters","Built for responsive agents"],"technologies":["reasoning","moe"],"openWeights":false,"architecture":{"type":"MoE","totalParams":"35B","activeParams":"3B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.59,"usageRank":26,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{"reasoningScore":32.4,"codingScore":18.2,"agentScore":2.7,"mathScore":25,"toolCallingScore":0.6},"benchmarksAsOf":{"reasoningScore":"2026-10-31","codingScore":"2026-10-31","agentScore":"2026-10-31","mathScore":"2026-10-31","toolCallingScore":"2026-10-31"},"popularity":20,"links":{"website":"https://www.upstage.ai"},"sources":["https://llm-stats.com/leaderboards/llm-leaderboard","https://openrouter.ai/rankings"],"providerName":"Upstage","blendedPrice":null},{"id":"pplx-decider-v1.1-27b","name":"Decider v1.1 27B","providerId":"perplexity","family":"Decider","kind":"decision","releaseDate":"2026-10-07","status":"ga","summary":"Perplexity’s open-weight multimodal decision model.","description":"Decider v1.1 27B is a new checkpoint of Perplexity’s open-weight multimodal decision model with the same API contract as v1. It reads text, JSON or images as state and returns typed decisions. Perplexity reports the top score on the Hugging Face Decision Index 0.3 (vendor-reported).","useCases":["extraction","high-volume","vision"],"strengths":["Open weights (Apache-2.0)","Image input"],"technologies":["system-one"],"openWeights":true,"license":"Apache-2.0","huggingFaceId":"perplexity-ai/pplx-decider-v1.1-27b","architecture":{"totalParams":"26B"},"modalities":{"input":["text","image"],"output":["decision"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"usage":{"tokenShare":0.01,"usageRank":68,"asOf":"2026-10-31","topTasks":[]},"benchmarks":{},"benchmarksAsOf":{},"popularity":20,"links":{"website":"https://www.perplexity.ai","huggingFace":"https://huggingface.co/perplexity-ai/pplx-decider-v1.1-27b"},"sources":["https://community.perplexity.ai/t/pplx-decider-v1-1-27b-our-updated-open-weights-multimodal-decision-model-is-now-available/6286","https://openrouter.ai/rankings"],"providerName":"Perplexity","blendedPrice":null},{"id":"mai-voice-2.1-flash","name":"MAI-Voice-2.1-Flash","providerId":"microsoft","family":"MAI-Voice","kind":"speech","releaseDate":"2026-10-01","status":"ga","summary":"Low-latency MAI-Voice for real-time, high-volume speech.","description":"MAI-Voice-2.1-Flash is the high-volume, latency-sensitive variant of MAI-Voice-2.1, with the same 23 languages. Microsoft reports 150 ms end-to-end latency for up to 45 seconds of audio (vendor-reported).","useCases":["realtime","high-volume","agents"],"strengths":["≈150 ms latency (vendor-reported)","23 languages"],"technologies":["realtime-voice"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["audio"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":20,"links":{"website":"https://microsoft.ai"},"sources":["https://microsoft.ai/news/our-first-streaming-transcription-model/"],"providerName":"Microsoft AI","blendedPrice":null},{"id":"mai-transcribe-2-streaming","name":"MAI-Transcribe-2-Streaming","providerId":"microsoft","family":"MAI-Transcribe","kind":"transcription","releaseDate":"2026-10-01","status":"ga","summary":"Microsoft AI’s first streaming speech-to-text model, in 60 languages.","description":"MAI-Transcribe-2-Streaming transcribes live audio in 60 languages with automatic, continuous language detection. Microsoft says it emits first partial text just over 100 ms into the audio, then revises it and commits stable text. It is offered in Microsoft Foundry, MAI Playground and Azure Voice Live.","useCases":["realtime","extraction","enterprise"],"strengths":["≈100 ms to first partial text (vendor-reported)","60 languages, auto-detected"],"technologies":["realtime-voice"],"openWeights":false,"architecture":{},"modalities":{"input":["audio"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":20,"links":{"website":"https://microsoft.ai"},"sources":["https://microsoft.ai/news/our-first-streaming-transcription-model/"],"providerName":"Microsoft AI","blendedPrice":null},{"id":"grok-voice-transcribe-2","name":"Grok Voice Transcribe 2.0","providerId":"xai","family":"Grok Voice","kind":"transcription","releaseDate":"2026-09-18","status":"ga","summary":"SpaceXAI’s speech-to-text model.","description":"Grok Voice Transcribe 2.0 is SpaceXAI’s speech-to-text model, offered through its API.","useCases":["extraction","high-volume"],"strengths":["SpaceXAI API"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["audio"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":20,"links":{"website":"https://x.ai"},"sources":[],"providerName":"xAI (SpaceXAI)","blendedPrice":null},{"id":"lfm-2.5-24b-a2b","name":"LFM2.5-24B-A2B","providerId":"liquid","family":"LFM","kind":"llm","releaseDate":"2026-10-13","status":"ga","summary":"Liquid’s largest LFM: a 24B MoE reasoning model with about 2.45B active parameters.","description":"LFM2.5-24B-A2B is Liquid AI’s largest LFM, a mixture of experts with 24B total and about 2.45B active parameters. It reasons on every request and is aimed at agents that call tools; weights are on Hugging Face.","useCases":["agents","reasoning","high-volume"],"strengths":["≈2.45B active parameters","Open weights","Built for tool calling"],"technologies":["reasoning","moe"],"openWeights":true,"huggingFaceId":"LiquidAI/LFM2.5-24B-A2B","architecture":{"type":"MoE","totalParams":"24B","activeParams":"2.45B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":18,"links":{"website":"https://www.liquid.ai","huggingFace":"https://huggingface.co/LiquidAI/LFM2.5-24B-A2B"},"sources":[],"providerName":"Liquid AI","blendedPrice":null},{"id":"clef-flash","name":"Clef-flash","providerId":"cloudflare","family":"Clef","kind":"decision","releaseDate":"2026-10-01","status":"ga","summary":"The fast 9B Clef for latency-critical, hot-path decisions.","description":"Clef-flash is the 9B member of Cloudflare’s open-source Clef decision family, fine-tuned from Qwen3.5-9B for latency-critical decisions. Cloudflare reports a 38.8 ms median latency and top scores on BFCL among Clef models (vendor-reported). Apache-2.0 weights on Hugging Face; served on Workers AI.","useCases":["extraction","realtime","high-volume"],"strengths":["≈39 ms median latency (vendor-reported)","Open weights (Apache-2.0)"],"technologies":["system-one","fine-tuning"],"openWeights":true,"license":"Apache-2.0","huggingFaceId":"Cloudflare/clef-flash","architecture":{"totalParams":"9B"},"modalities":{"input":["text","image"],"output":["decision"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":18,"links":{"website":"https://developers.cloudflare.com/workers-ai/","huggingFace":"https://huggingface.co/Cloudflare/clef-flash"},"sources":["https://developers.cloudflare.com/changelog/post/2026-10-01-clef-workers-ai/"],"providerName":"Cloudflare","blendedPrice":null},{"id":"lfm-2.5-embedding-350m","name":"LFM2.5-Embedding-350M","providerId":"liquid","family":"LFM 2.5","kind":"embedding","releaseDate":"2026-08-18","status":"ga","summary":"Tiny open embedding model for on-device semantic search.","description":"LFM2.5-Embedding-350M is a compact text embedding model from Liquid AI producing 1,024-dimension vectors for retrieval and semantic search; small enough to run on edge devices.","useCases":["on-device","high-volume"],"strengths":["350M parameters","Open weights"],"technologies":["vector-search"],"openWeights":true,"huggingFaceId":"LiquidAI/LFM2.5-Embedding-350M","architecture":{"totalParams":"350M"},"modalities":{"input":["text"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":18,"links":{"website":"https://www.liquid.ai","huggingFace":"https://huggingface.co/LiquidAI/LFM2.5-Embedding-350M"},"sources":[],"providerName":"Liquid AI","blendedPrice":null},{"id":"pplx-embed-v2-late-9b","name":"pplx-embed-v2-late-9b","providerId":"perplexity","family":"pplx-embed v2","kind":"embedding","releaseDate":"2026-10-07","status":"ga","summary":"Perplexity’s large late-interaction embedder for text and visual document retrieval.","description":"pplx-embed-v2-late-9b is the large late-interaction embedding model from Perplexity, producing 128-dimension per-token vectors for text and images. Perplexity reports the highest average on its domain-specific text retrieval suite (81.3% nDCG@10) and 64.0% on BrowseComp+ (vendor-reported).","useCases":["research","vision","enterprise"],"strengths":["Searches PDF pages without OCR","Multi-vector (ColBERT-style) scoring"],"technologies":["vector-search","rag"],"openWeights":true,"license":"MIT","huggingFaceId":"perplexity-ai/pplx-embed-v2-late-9b","architecture":{"totalParams":"8.4B"},"modalities":{"input":["text","image"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":18,"links":{"website":"https://www.perplexity.ai","huggingFace":"https://huggingface.co/perplexity-ai/pplx-embed-v2-late-9b"},"sources":["https://www.perplexity.ai/hub/blog/multimodal-embeddings-beyond-a-single-vector"],"providerName":"Perplexity","blendedPrice":null},{"id":"kandinsky-6-video","name":"Kandinsky 6.0 Video","providerId":"kandinsky-lab","family":"Kandinsky","kind":"video","releaseDate":"2026-10-06","status":"ga","summary":"MIT-licensed open video model that generates clips with synchronized audio.","description":"Kandinsky 6.0 Video is Sber’s open-weight video family: a 29B Pro and a 3B Lite tier, each in pretrained and distilled (fewer-step) variants, plus a separate super-resolution model that upscales to Full HD. It generates about 5-second clips with synchronized 44 kHz audio from text or an image. Code and weights are MIT-licensed.","useCases":["multimodal","on-device"],"strengths":["Open weights (MIT)","Synchronized audio","Pro and Lite tiers"],"technologies":["diffusion"],"openWeights":true,"license":"MIT","huggingFaceId":"kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers","architecture":{"totalParams":"29B"},"modalities":{"input":["text","image"],"output":["video","audio"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":18,"links":{"website":"https://github.com/kandinskylab","huggingFace":"https://huggingface.co/kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers"},"sources":["https://github.com/kandinskylab/kandinsky-6"],"providerName":"Kandinsky Lab (Sber)","blendedPrice":null},{"id":"mellum-2.1","name":"Mellum2.1","providerId":"jetbrains","family":"Mellum","kind":"llm","releaseDate":"2026-10-08","status":"ga","summary":"JetBrains’ open 12B MoE (2.5B active) built to work inside coding agents.","description":"Mellum2.1 is an open-weight 12B mixture-of-experts model (2.5B active) from JetBrains. The architecture is unchanged from Mellum2; gains come from post-training, mainly reinforcement learning. It is aimed at sub-agent work such as diagnosing failing tests and drafting fixes, and at private self-hosted deployment. JetBrains reports about 1.6× faster single-request decoding with its multi-token-prediction head on one H200 (vendor-reported).","useCases":["coding","agents","on-device"],"strengths":["Open weights (Apache-2.0)","2.5B active parameters","MTP head for fast decoding"],"technologies":["speculative-decoding","reasoning","moe"],"openWeights":true,"license":"Apache-2.0","huggingFaceId":"JetBrains/Mellum2.1-12B-A2.5B-Thinking","architecture":{"type":"MoE","totalParams":"12B","activeParams":"2.5B"},"modalities":{"input":["text"],"output":["text"]},"features":{"reasoning":true,"toolCalling":true,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":16,"links":{"website":"https://www.jetbrains.com","huggingFace":"https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking"},"sources":["https://blog.jetbrains.com/ai/2026/10/mellum2-1-gets-to-work-a-fast-open-model-for-coding-agents/"],"providerName":"JetBrains","blendedPrice":null},{"id":"d1-3b","name":"d1-3B","providerId":"liquid","family":"d1","kind":"decision","releaseDate":"2026-10-07","status":"ga","summary":"Liquid’s open 3B decision model, fast enough for robots and edge devices.","description":"d1-3B is an open-weight Liquid AI decision model built on LFM2.5-VL-3B. It answers typed questions about text and images in a single forward pass instead of generating tokens. Liquid reports 48.57 on Decision Index v0.2.1 (best under 10B) and 8 ms per decision on an RTX 4090 (vendor-reported).","useCases":["extraction","realtime","on-device","vision"],"strengths":["Runs on Jetson-class hardware","Image input","Open weights"],"technologies":["system-one"],"openWeights":true,"huggingFaceId":"LiquidAI/d1-3B","architecture":{"totalParams":"3B"},"modalities":{"input":["text","image"],"output":["decision"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":16,"links":{"website":"https://www.liquid.ai","huggingFace":"https://huggingface.co/LiquidAI/d1-3B"},"sources":["https://www.liquid.ai/blog/d1-open"],"providerName":"Liquid AI","blendedPrice":null},{"id":"pplx-embed-v2-late-0.6b","name":"pplx-embed-v2-late-0.6b","providerId":"perplexity","family":"pplx-embed v2","kind":"embedding","releaseDate":"2026-10-07","status":"ga","summary":"Small late-interaction (ColBERT-style) embedder that searches text and PDF pages.","description":"pplx-embed-v2-late-0.6b is a late-interaction embedding model from Perplexity: it keeps 128-dimension vectors per token and scores with MaxSim instead of one vector per input. It embeds text and images, so rendered PDF pages can be searched without OCR, and shares an embedding space with the 9B model so it can query a 9B-built index. Distilled from an 18B teacher (594M parameters).","useCases":["research","vision","high-volume"],"strengths":["Searches PDF pages without OCR","Shares an index with the 9B model"],"technologies":["vector-search","rag"],"openWeights":true,"license":"MIT","huggingFaceId":"perplexity-ai/pplx-embed-v2-late-0.6b","architecture":{"totalParams":"594M"},"modalities":{"input":["text","image"],"output":["embedding"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":16,"links":{"website":"https://www.perplexity.ai","huggingFace":"https://huggingface.co/perplexity-ai/pplx-embed-v2-late-0.6b"},"sources":["https://www.perplexity.ai/hub/blog/multimodal-embeddings-beyond-a-single-vector"],"providerName":"Perplexity","blendedPrice":null},{"id":"stepaudio-3-asr","name":"StepAudio 3 ASR","providerId":"stepfun","family":"StepAudio","kind":"transcription","releaseDate":"2026-09-14","status":"ga","summary":"StepFun’s speech recognition model.","description":"StepAudio 3 ASR is StepFun’s dedicated speech-to-text model.","useCases":["extraction","multimodal"],"strengths":["Dedicated ASR model"],"technologies":[],"openWeights":false,"architecture":{},"modalities":{"input":["audio"],"output":["text"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":16,"links":{"website":"https://www.stepfun.com"},"sources":[],"providerName":"StepFun","blendedPrice":null},{"id":"clef-omni","name":"Clef Omni","providerId":"cloudflare","family":"Clef","kind":"decision","releaseDate":"2026-10-09","status":"ga","summary":"Clef’s mixture-of-experts member, deciding over text, images, audio and video.","description":"Clef Omni is the mixture-of-experts model in Cloudflare’s open-source Clef decision family. Cloudflare fine-tuned it from Qwen3-Omni-30B-A3B (3B of 30B parameters active) and serves it on Workers AI. It takes text, images, audio or video as the state and answers typed questions with probabilities.","useCases":["extraction","multimodal","realtime"],"strengths":["Audio and video input","Open weights (Apache-2.0)","3B active parameters"],"technologies":["system-one","fine-tuning","native-multimodal","moe"],"openWeights":true,"license":"Apache-2.0","huggingFaceId":"Cloudflare/clef-omni","architecture":{"type":"MoE","totalParams":"30B","activeParams":"3B"},"modalities":{"input":["text","image","audio","video"],"output":["decision"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":14,"links":{"website":"https://developers.cloudflare.com/workers-ai/","huggingFace":"https://huggingface.co/Cloudflare/clef-omni"},"sources":[],"providerName":"Cloudflare","blendedPrice":null},{"id":"mercury-decide","name":"Mercury Decide","providerId":"inception","family":"Mercury","kind":"decision","releaseDate":"2026-09-30","status":"ga","summary":"Inception’s structured decision endpoint: choices, scores and yes/no with probabilities.","description":"Mercury Decide is Inception’s decision model, offered as a System One endpoint. You pass in text or images plus typed questions, and it answers each with a choice, a score or a yes/no, along with probabilities.","useCases":["extraction","realtime","high-volume"],"strengths":["Image input","Choice, score and yes/no answers"],"technologies":["system-one"],"openWeights":false,"architecture":{},"modalities":{"input":["text","image"],"output":["decision"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":12,"links":{"website":"https://www.inceptionlabs.ai"},"sources":[],"providerName":"Inception","blendedPrice":null},{"id":"helix-world-preview","name":"Helix World (preview)","providerId":"noiz","family":"Helix World","kind":"world-model","releaseDate":"2026-10-12","status":"preview","summary":"Real-time interactive world model that generates video and spatial audio together.","description":"Helix World, from Noiz AI and HKUST, generates an explorable video scene and matching spatial stereo audio from a single starting image (with an optional text prompt), steered in real time by camera motion and WASD/arrow-key input. It renders 768×512 at 24 fps in ~1.3-second chunks with a real-time factor of 0.77 on one H800 GPU, including decoding. Built on the open LTX-2.3 audio-visual model; the public checkpoint is a preview.","useCases":["multimodal","research","realtime"],"strengths":["Video + spatial stereo audio","Faster than real time on one GPU","Open weights"],"technologies":["diffusion"],"openWeights":true,"license":"Apache-2.0","huggingFaceId":"NoizAI/HelixWorld-preview","architecture":{},"modalities":{"input":["text","image"],"output":["video","audio"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":false,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":12,"links":{"website":"https://helixworld.org","huggingFace":"https://huggingface.co/NoizAI/HelixWorld-preview"},"sources":["https://noiz.ai/lp/blog/meet-helix-world","https://arxiv.org/abs/2609.38123"],"providerName":"Noiz AI","blendedPrice":null},{"id":"solar-decide","name":"Solar Decide","providerId":"upstage","family":"Solar","kind":"decision","releaseDate":"2026-09-28","status":"ga","summary":"Upstage’s structured decision endpoint built on Solar Mini 4.","description":"Solar Decide is Upstage’s decision model, offered as a System One endpoint running on Solar Mini 4. You pass in a state plus typed questions, and it answers each with a choice, a score or a yes/no, along with probabilities. Upstage also offers a lower-latency Flash variant.","useCases":["extraction","long-context","enterprise"],"strengths":["Built on Solar Mini 4","Flash variant for lower latency"],"technologies":["system-one"],"openWeights":false,"architecture":{},"modalities":{"input":["text"],"output":["decision"]},"features":{"reasoning":false,"toolCalling":false,"structuredOutput":true,"batchApi":false},"pricing":{"input":null,"output":null},"benchmarks":{},"benchmarksAsOf":{},"popularity":10,"links":{"website":"https://www.upstage.ai"},"sources":[],"providerName":"Upstage","blendedPrice":null}],"total":149}