Google released EmbeddingGemma 2 on October 6, 2026: a 740-million-parameter open embedding model that maps text, code, images, audio and video into one shared vector space. It is released under the Apache 2.0 license, is designed to run on phones and laptops, and has an 8K-token context window, four times that of the original EmbeddingGemma. Among the September–October 2026 releases, it is the open-weight model aimed squarely at search and retrieval.
An embedding model turns content into vectors so that similar items sit close together; it is the retrieval half of most search and RAG systems. Google's announcement says the original text-only EmbeddingGemma passed 20 million downloads, and that version 2 extends it to code, images, audio and video, built on the Gemma 4 architecture and the same technology as Google's Gemini Embedding models.
The details that matter for engineering, from the announcement and the model card:
The announcement and model card both state Apache 2.0, a permissive open-source license that allows commercial use. For an embedding model that may end up inside a shipped product or a customer's device, a standard license removes a review step.
Google says EmbeddingGemma 2 leads sub-1B multimodal embedding models on benchmarks including MTEB Code and MAEB (audio), and matches or beats many larger models. The model card's most concrete figure is code retrieval: MTEB Code rises from 68.76 to 78.68 against EmbeddingGemma 1, while multilingual text performance stays roughly level (61.36 against 61.15 on multilingual MTEB). These are Google's reported results, and this article does not reproduce them.
The pattern is useful in itself. If your workload is multilingual text search, version 2 is not a big quality jump; its value is the new modalities, the longer context and code retrieval.
Changing embedding models means re-embedding. Vectors produced by different models are not comparable, so moving from EmbeddingGemma 1 or any other embedder to version 2 requires re-indexing the stored corpus. Budget the compute and a cut-over plan.
Pick dimensions deliberately. Truncating to 256 or 128 dimensions saves storage and speeds search, at some cost in retrieval quality. Test on your own queries before choosing.
Use the modular encoders. A text-and-code search service can skip the vision and audio encoders entirely and run at the smaller footprint.
One space for many media. Because text, images, audio and video share a vector space, a text query can retrieve a video moment or an audio clip without separate indexes per media type. That simplifies cross-modal search, but evaluate retrieval quality per modality rather than assuming it is uniform.
Teams running local or private RAG that want embeddings to stay on device or on their own servers.
Products with media libraries, such as photo, video, voice memo or meeting archives, that want text-to-media search from one model.
Code search and coding-agent retrieval, where Google reports the largest quality gain.
For the models that consume what these embeddings retrieve, see the other open releases in this series, including DeepSeek-V4.1-Flash and Holo4.
When was EmbeddingGemma 2 released?
October 6, 2026.
What license does EmbeddingGemma 2 use?
Apache 2.0, per Google's announcement and the model card.
How many dimensions does EmbeddingGemma 2 output?
768 by default, with Matryoshka truncation to 512, 256 or 128.
Can EmbeddingGemma 2 run on a phone?
Google reports about 191MB of active RAM for text-only use and about 567MB for the full multimodal model on a Pixel 11 Pro with quantization.
Discover more content:
