Small, multimodal, and permissively licensed
Google released EmbeddingGemma 2 on 6 October, an open embedding model that maps text, code, images, video and audio into a single shared vector space. The full multimodal model is 740 million parameters. The licence is Apache 2.0, which means commercial use without negotiation.
The design is modular rather than monolithic. Text-only workloads can run a 270-million-parameter configuration. Quantised, Google puts active memory at roughly 191MB for text-only and about 567MB for the full multimodal model — small enough to sit on a phone rather than behind an API call.
What the numbers say, and who measured them
The headline benchmark Google published is MTEB Code, where it reports 78.68, a 9.92-point improvement over the first EmbeddingGemma. Google also says the model achieves leading scores among sub-1B multimodal embedders across MTEB and MAEB, and matches or outperforms larger models on text, vision and audio tasks.
Those are Google’s own reported figures on public benchmarks. The useful property of open weights is that they are checkable: anyone can download the model and re-run MTEB themselves, which is not true of the frontier numbers published in the same week.
Storage is the other practical axis. Vectors truncate to 768, 512, 256 or 128 dimensions, which Google says allows up to a sixfold reduction in storage. For a retrieval index on a device, that is the difference between fitting and not fitting.
Why an embedding release matters more than it sounds
Embeddings are the unglamorous layer underneath retrieval, recommendation, deduplication and semantic search. A single model that puts a photo, a voice note, a video clip and a paragraph into the same space lets a device search across all of them with one index rather than four.
Doing that locally is the point. An on-device embedder means a phone can build and query a semantic index over a user’s own files without sending any of them to a server — the architecture that privacy regulators have been pushing toward and that the EU’s transparency rules make cheaper to defend.
Google says the model is available on Hugging Face and Kaggle, with optimised builds on LiteRT Community, and that it is coming to the Gemini Enterprise Agent Platform Model Garden. The thing to watch is whether independent MTEB and MAEB runs reproduce the sub-1B leadership claim once the community has had the weights for a few weeks.
