EmbeddingGemma 2
Open 740M multimodal embedding model that runs on phones.
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).
- Provider
- Type
- Embedding model
- Released
- Oct 6, 2026
- Input
- text, image, audio, video
- Output
- embedding
- Open weights
- Yes
- License
- Apache-2.0
Best for
- Local & on-device
- Multimodal
- Coding
Strengths
- Open weights (Apache-2.0)
- Text, image, audio and video
- Runs on-device
More from Google
- Gemini 4 Argon — Google. Google’s flagship reasoning model for software engineering, science and long-running agents.
- Gemini 3.8 Flash — Google. Google’s most capable Flash model, with text, image, audio and video input.
- Gemini 3.5 Flash Lite — Google. Google’s efficient multimodal model for focused subagent tasks.
- Nano Banana 2.1 — Google. Google’s latest image generation and editing model.
- Gemma 4 31B — Google. Open-weight dense multimodal model from Google DeepMind.
- Gemma 4 26B A4B — Google. Open MoE Gemma with only 3.8B active parameters.
- Gemini Embedding 2 — Google. Google’s first multimodal embedding model.
- Gemini Omni Flash 1.1 — Google. Google’s fast video generation model.