EmbeddingGemma 2: 740M-param open model goes multimodal

EmbeddingGemma arrived last year as a lightweight option for high-quality text embeddings, aimed at helping apps organize, search, and connect information directly on consumer hardware. Google DeepMind says the developer community's response blew past its expectations, with more than 20 million downloads and use in on-device search tools and privacy-first retrieval pipelines. EmbeddingGemma 2 is the follow-up, built on the Gemma 4 architecture and expanding beyond text to unify code, images, video, and audio in a shared embedding space.

The design choices point in one direction: staying small enough to run locally. It has 740 million parameters, but can need as little as 270M parameters for text-only workloads, with optional vision (170M) and audio (300M) encoders. Matryoshka Representation Learning lets developers truncate output vectors from 768 dimensions down to 512, 256, or 128, which Google says provides up to 6x storage reduction for local vector databases. Its 8K token context window, 4x larger than EmbeddingGemma 1, lets it process up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations.

On quality, Google says EmbeddingGemma 2 leads among sub-1B multimodal embedders on benchmarks like MTEB Code and MAEB, and matches or outperforms many larger models across text, vision, and audio. The outcome hinges on whether that quality-per-parameter claim holds up as developers test it on their own hardware, and on whether the memory figures translate into smooth performance outside Google's own Pixel 11 Pro. Google says Gemini Enterprise Agent Platform Model Garden availability is coming soon.