Google DeepMind releases EmbeddingGemma 2, a small open model that puts text, code, images, video and audio in one search space
EmbeddingGemma 2 has 740 million parameters, runs on phones and is licensed under Apache 2.0. It lets apps find, for example, a video clip from a voice memo — locally, without sending data to the cloud.
Google DeepMind launched EmbeddingGemma 2 on 6 October 2026, an open embedding model — the kind of model that turns content into vectors so apps can search and match it. The first EmbeddingGemma (2025, text only) has passed 20 million downloads, Google says.
What is new
- Five kinds of input in one space: text, code, images, video and audio are mapped into a shared 768-dimensional vector space, so you can, for example, find a video clip with a voice memo or search hours of audio with a text query.
- Small and modular: 740 million parameters in total — a 270M text model plus a 170M vision encoder and a 300M audio encoder. Developers can load only the parts they need (270M for text and code only).
- Built on the Gemma 4 architecture, released under the commercially permissive Apache 2.0 licence; weights on Hugging Face and Kaggle.
- Context window of 8,192 tokens, four times that of version 1. Vectors can be shortened to 512, 256 or 128 dimensions to save storage.
- Google reports the MTEB Code score rose from 68.76 to 78.68, and claims leading results among multimodal embedders under one billion parameters.
- With quantisation on a Pixel 11 Pro, the text-only model needs about 191 MB of active RAM and the full multimodal model about 567 MB, according to Google.
Limitations stated by Google
The model card says EmbeddingGemma 2 has not undergone post-training safety alignment; performance may differ across its 100+ languages; and it should run in bfloat16 or float32, because float16 can produce broken embeddings. Benchmark figures are Google’s own.
Also available in: Саха тылаРусскийEspañol中文Portuguêsالعربية