Google DeepMind open-sources EmbeddingGemma 2: a 740M-parameter multimodal embedding model that puts text, code, images, audio and video in one vector space and runs in under 600MB on a phone

Why it matters

Key points

Google DeepMind released EmbeddingGemma 2 on October 6, 2026 (19:57 GMT). It is the second generation of last year's text-only EmbeddingGemma, which Google says has passed 20 million downloads, and it now maps code, images, video and audio into the same embedding space as text.

The model is built on the Gemma 4 architecture, has 740 million parameters and ships under the commercially permissive Apache 2.0 license. It is modular: text-only use needs as little as 270M parameters, with optional vision (170M) and audio (300M) encoders. Using Matryoshka Representation Learning, output vectors can be truncated from 768 to 512, 256 or 128 dimensions, cutting local vector storage by up to six times. The context window is 8K tokens, four times the first version, enough for about 5.5 minutes of audio, 29 images or 58 video frames. Quantized on a Pixel 11 Pro, it uses about 191MB of active RAM for text and about 567MB for the full multimodal model.

Google says multilingual text quality matches the first version, code retrieval on MTEB Code rises from 68.76 to 78.68 (about 9.9 points), and the model leads sub-1B multimodal embedders on benchmarks including MTEB Code and MAEB (audio). Weights are on Hugging Face and Kaggle, with support in transformers, sentence-transformers, vLLM, llama.cpp, Ollama, MLX and others, plus demos in Google AI Edge Gallery.

Verifiable

Primary page: https://deepmind.google/blog/embeddinggemma-2-an-open-lightweight-multimodal-embedding-model/ (Google DeepMind blog RSS pubDate Tue, 06 Oct 2026 19:57 GMT; authors Sahil Dua and Henrique Schechter Vera, research engineers at Google DeepMind). Parameter counts, memory figures and the MTEB Code score come from the official post, with full evaluations in the model card. The best-in-class claim is Google's own comparison and has not yet been independently re-run; the phone memory numbers are for quantized weights on a single device.

Our take

The interesting part is not another small open model but retrieval moving onto the phone. Multimodal search has usually meant uploading photos, recordings and video to the cloud and processing each separately. A model that runs in under 600MB and puts text, audio and video in one vector space makes things like finding a video clip on your phone from a spoken query, or searching hours of recordings offline, practical, and sidesteps privacy and latency problems. Sharing a tokenizer and audio encoder with Gemma 4 shows Google is pushing a complete on-device retrieval-plus-generation stack rather than a single model.

The 9.9-point code retrieval gain matters most for coding agents, which can index a whole codebase locally instead of calling a cloud embedding service. Still, this is retrieval infrastructure and does not move the frontier of model capability, so it goes to Latest rather than this week's major changes. Two things to watch: independent results on public leaderboards such as MTEB, and whether the open-source community adopts it as the default for on-device RAG the way it did the first version.