EmbeddingGemma 2 is an open multimodal embedding model from Google DeepMind, built for developers and startups who want privacy-first, on-device search across text, code, images, video, and audio. Google released it on October 6, 2026, and it is the first widely available embedding model that natively handles five modalities in one space.
Core Features
- One shared space: 740 million parameters map text, code, images, video, and audio into a single 768-dimensional vector space for cross-modal retrieval.
- Modular encoders: load only what you need, from a 270M text-only build up to the full 740M multimodal model.
- Matryoshka truncation: shrink the 768-dim output to 512, 256, or 128 dimensions for up to 6x storage savings with minimal accuracy loss.
- Longer context: an 8,192-token window, four times the first version, covering about 29 images or 5.5 minutes of audio.
- Apache-2.0 weights on Hugging Face, Kaggle, Ollama, and llama.cpp, with commercial use allowed.
Use Cases / Best For
- Multimodal RAG that indexes code, screenshots, meeting audio, and video clips without a remote API.
- On-device apps where raw media must stay local for privacy.
- Teams wanting one embedding model instead of separate text and image services.
Pros / Cons
- Pro: One open model covers text, code, image, video, and audio.
- Pro: Runs on-device, keeping raw media local for privacy.
- Con: You host and operate it; there is no managed API.
- Con: The full multimodal build is 740M params, heavier than text-only options.
Pricing
EmbeddingGemma 2 is free and open under Apache-2.0 ($0). There is no paid API. It runs on a MacBook Air at about 47 milliseconds per sentence on CPU, and the predecessor passed 20 million downloads. On MTEB Code it improved from 68.76 to 78.68.
Our Take: Best for builders who need a single open, on-device embedding model across media types; the trade-off is that you host and operate it yourself, with no managed endpoint.
Explore more AI Search tools.
