Google releases EmbeddingGemma 2, a 740M multimodal embedding model
Apache 2.0 model unifies text, image, audio and video for on-device search and RAG
The facts
- EmbeddingGemma 2 has 740 million parameters, is built on the Gemma 4 architecture, and is released under Apache 2.0
- Google says the model requires as little as ~191MB active RAM (text-only, quantised) and ~567MB for the full multimodal model on a Google Pixel 11 Pro
- It has an 8K token context window, 4x larger than EmbeddingGemma 1, handling up to 5.5 minutes of audio, 29 images or 58 video frames
- Google reports a 9.92-point improvement on MTEB Code (from 68.76 to 78.68) versus the original EmbeddingGemma, and says it is 'best-in-class' among sub-1B multimodal embedders on MTEB Code and MAEB
- Matryoshka Representation Learning lets developers truncate output vectors from 768 dimensions down to 128, which Google says gives up to 6x storage reduction
- The model is modular: as little as 270M parameters for text-only, with optional 170M vision and 300M audio encoders
- Weights are available on Hugging Face and Kaggle, with Gemini Enterprise Agent Platform Model Garden availability coming soon; the predecessor model has had over 20 million downloads
Google DeepMind has released EmbeddingGemma 2, a 740-million-parameter embedding model. It maps text, images, audio and video into a single shared embedding space. It is small enough to run on a phone.
The model is built on the Gemma 4 architecture. It ships under the commercially permissive Apache 2.0 licence. It succeeds the text-only EmbeddingGemma, which Google says has passed 20 million downloads since last year.
The model is designed to run entirely on-device. Google says that with quantisation it needs as little as ~191MB of active RAM for text-only inference, and ~567MB for the full multimodal version, measured on a Google Pixel 11 Pro. It is modular. A 270M-parameter core handles text alone, with optional 170M vision and 300M audio encoders bolted on for full multimodal support. Its context window has grown fourfold to 8,000 tokens. Google says that is enough to process 5.5 minutes of audio, 29 images or 58 video frames in a single pass.
On quality, Google reports the model leads sub-1B multimodal embedders on the MTEB Code and MAEB benchmarks. It says the model matches or beats some larger specialist models. Code retrieval saw the largest stated jump. Google's own MTEB Code score rose from 68.76 to 78.68, a gain it attributes to making the model useful for local codebase indexing and coding-agent retrieval. Matryoshka Representation Learning lets developers truncate the 768-dimension output vectors down to 128 dimensions. Google says this cuts storage for local vector databases by up to 6x.
Google built the model to share a tokenizer and audio encoder with Gemma 4. The two can run together in a single pipeline, for on-device retrieval-augmented generation with a lower combined memory footprint. Weights are available now on Hugging Face and Kaggle. Support is already in place for MediaPipe, LiteRT, transformers.js, vLLM, llama.cpp, SGLang, Ollama and LMStudio. A Gemini Enterprise Agent Platform Model Garden listing is coming soon.
Why it matters
The sub-gigabyte footprint and shared tokenizer with Gemma 4 let developers build fully offline, cross-modal RAG on a phone. Search across photos, voice memos and video need never touch a server. That puts pressure on cloud embedding APIs wherever privacy or latency matters more than raw accuracy.
Primary source
Google DeepMind — deepmind.googleWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.