Google unveiled EmbeddingGemma 2 this week, a unified embedding model that handles text, code, image, video, and audio retrieval through a single 740-million-parameter architecture. Testing on a Pixel 11 Pro showed the model consuming approximately 567MB of active RAM when quantized. The approach allows developers to search across multiple data types without preprocessing—images need no captions, audio requires no transcription—all mapped into a shared 768-dimensional vector space.

The model weights are available under Apache 2.0 licensing. Deployment on-device is accessible now through LiteRT and MediaPipe Tasks, with Android ML Kit integration and NPU acceleration arriving in the coming weeks. The foundation builds on Gemma 4's architecture.

Modular encoders, one vector space

Developers load only the encoders their application needs. The text and code encoder operates at 270 million parameters, consuming roughly 191MB of active RAM on the test device. Adding the vision encoder for images and video brings the model to 440 million parameters; the audio encoder alone reaches 570 million; both together total the full 740 million. All configurations project into the same embedding space, enabling teams to expand from text-only indexes to include image or audio search without re-embedding previously stored data.

Google also quadrupled the context window from 2,048 to 8,192 tokens, which the company says covers as much as 5.5 minutes of audio, 29 images or 58 video frames in a single input.

Video sampling defaults to one frame per second, meaning those 58 frames represent just under a minute of footage. Google's Video Moments Finder demonstration indexes video frames and audio segments locally, then searches them using plain text queries to pinpoint specific moments without generating captions or transcripts. The Instant Media Search feature applies the same technique to phone photos and videos, storing embeddings in SQLite and refreshing results as the user types.

Matryoshka shrinks the index

On mobile devices, vector indexes can rival the model itself in storage footprint—a million 768-dimensional bfloat16 vectors occupy roughly 1.5GB. EmbeddingGemma 2 incorporates Matryoshka Representation Learning, allowing developers to reduce embeddings to 512, 256, or 128 dimensions without retraining. At 256 dimensions, a million-vector index shrinks to approximately 500MB.

Google says the shorter embeddings keep most of their full-size quality on text and code and roughly 95% on image, video and speech retrieval.

The trade-off becomes more pronounced at 128 dimensions, where Google reports text and code retain around 90% quality but multimodal retrieval drops to approximately 75%. The company recommends testing this configuration against real data before deploying it for multimodal queries. This efficiency-versus-quality balance reflects a broader trend in embedding releases this season, as seen when Cohere's optimized query model showed minimal retrieval degradation in its benchmarks.

On-device code search

Code uses the same 270-million-parameter base as text. Google reports an MTEB Code score of 78.68 for EmbeddingGemma 2, a significant jump from 68.76 for the original EmbeddingGemma. In a demonstration combining code retrieval with agent workflows, Google indexed the Hugging Face Transformers repository using the text-only configuration and paired it with Gemma 4 26B A4B running in Pi. The embedding model handled repository retrieval while the larger model orchestrated the agent, contrasting with a previous workaround that consumed 18,000 tokens before performing any work.

The embedding model also supports classification through MediaPipe Decision, which compares incoming embeddings against candidate descriptions rather than generating responses. A chess demonstration showed the system evaluating 500 options per turn in under 100 milliseconds. This approach offers a cheaper alternative for decisions that don't require generated answers, addressing a common inefficiency where agents exhaust tokens on choices that could be resolved through comparison alone.

Leaner local RAG pipelines

EmbeddingGemma 2 adopts Gemma 4's text tokenizer and audio encoder design, reducing memory overhead when both models run together on the same device. Google has demonstrated multimodal retrieval functioning on its flagship Pixel device. Real-world performance with larger indexes, diverse hardware platforms, and non-Google applications remains to be seen.