Background
EmbeddingGemma was Google's text-only embedding model. An embedding model converts content into vectors so that software can measure how similar two pieces of content are, which is the basis of search and retrieval features. The first version handled text, and the release extended that to other kinds of input.
What happened
Google published its blog post on EmbeddingGemma 2 on 17 November 2026. The post links a technical report that describes the work.[1] The release extends the text-only EmbeddingGemma with native multimodal understanding.[1] Google says the recipe builds on Gemma 4 and Gemini Embedding 2.[1]
Key details
Google's model card describes the model as 740M parameters, with a 768-dimensional unified space for its inputs.[2]
On the MTEB code benchmark, Google reports 78.68 for EmbeddingGemma 2 against 68.76 for EmbeddingGemma 1.[2] MTEB is a public suite of tests for embedding models, and its code tasks measure how well a model places code close to matching descriptions. The comparison is Google's own report, so it is best read as a before-and-after measurement from the same source.
The model card lists Apache 2.0 licensing.[2]
Impact
For developers, the release adds a single open model that covers several kinds of content, instead of one model per type. The code score is the figure most relevant to people who search codebases or documentation with embeddings, though the right choice for any project depends on its own data and tests.
Sources
- 1
- 2
Cite this page: Google releases EmbeddingGemma 2 with multimodal embeddings. TellWorld, https://www.tellworld.net/news/embeddinggemma-2-2026. Updated 2026-11-18