Google has released EmbeddingGemma 2, expanding its open embedding model beyond text to images, audio and video. The model maps those formats into one embedding space, a useful step for systems that need to search or compare mixed media rather than treat every format as a separate island.
Google DeepMind Watch analysis
What happened
The EmbeddingGemma 2 team announced the release on 17 November. Its model can process text, images, audio and video, including combinations of them, and maps them into a shared embedding space. Google also says it has made the model backbone more compact.
The team has published a technical report covering its training strategy, evaluation methodology and experimental results. That gives developers and researchers material to assess beyond the announcement’s headline claims.
Why it matters
Embeddings help software represent information in a form that can be compared or retrieved. A shared space for text, images, audio and video could make it easier to build search and retrieval systems across different media, without stitching together separate representations for each format.
A more compact backbone may also matter to teams choosing what they can afford to run. The supplied announcement does not give the model’s size, benchmark results or deployment requirements, so those practical trade-offs need to be judged from the technical report rather than inferred from the word “compact”.
Our read
The notable change is not simply another model with a new number: EmbeddingGemma 2 brings multiple media types into one embedding space. That is a concrete capability for developers to evaluate. The technical report is where the interesting questions begin: whether the model performs well across formats, and what its compactness costs or saves in practice.
What to watch
- How the report’s evaluations compare with specialist models and earlier embedding systems.
- What hardware and memory are needed to run the model.
- Whether developers build useful search or retrieval tools around its multimodal embeddings.
Discussion spark: For multimodal search, would you choose one shared embedding model for every format, or separate specialist models if they perform better?
Sources and evidence
- Read the EmbeddingGemma 2 technical report. – blog.google (17 November 2026, 17:03 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind