Google's EmbeddingGemma 2 embeds text, images and audio on-device
A 740M-parameter open model maps five modalities into one vector space, runs in under 600MB of RAM, and ships under Apache 2.0.
Google DeepMind released EmbeddingGemma 2 on 6 October, an open embedding model that maps five modalities — text, code, images, video and audio — into a single 768-dimensional vector space. It’s modular: a 270M-parameter text encoder with optional 170M vision and 300M audio encoders, built on Gemma 4, for a 740M-parameter total.
The context window is 8K tokens, four times the original EmbeddingGemma. On MTEB Code it scores 78.68, nearly 10 points above version 1, and Google says it beats specialist models twice its size on image, video, document and audio retrieval. Matryoshka training lets the 768-dim output truncate to 512, 256 or 128 dimensions for smaller indexes. Quantised, the full multimodal model needs about 567MB of RAM on a Pixel 11 Pro; text-only weights need about 191MB. It’s Apache 2.0, downloadable from Hugging Face and Kaggle, and already supported in llama.cpp, Ollama, LM Studio, MLX, transformers.js and vLLM.
Why it matters: this is a small, permissively licensed model for semantic search across mixed media with no server round-trip — a voice memo can retrieve a video clip, or text can search hours of audio, entirely on-device. Anyone building retrieval or RAG over non-text content gets a serious alternative to running separate embedding models per modality.
The caveat: the benchmark numbers are Google’s own; there’s no independent MTEB leaderboard entry yet to check them against.