Cloudflare announced on November 12, 2026 that Google DeepMind’s EmbeddingGemma 2 is available through Workers AI. The practical change is a hosted route to searching mixed collections of text, pictures, audio and video, rather than a new chatbot that writes answers.
Why a shared search space matters
An embedding is a numerical representation used to compare content. Putting different media in a compatible space can help a text query find an image or a recording, not only a document containing the same words. For an IT team, that could make a knowledge collection easier to search without building a separate retrieval path for every format.
Cloudflare lists a default output of 768 dimensions and shorter 128, 256 and 512 alternatives. Its dated availability announcement is the news here; it is not evidence that the model was first created that day or that every application benefits from switching.
Three checks before changing an index
- Keep a consistent representation. Queries and stored content must use compatible embeddings and the same chosen size. Plan a separate test index rather than mixing old and new vectors and assuming the rankings remain meaningful.
- Test the media you actually have. Include difficult searches, irrelevant results and mixed Hebrew-English material. A shared vector space is a capability, not a measured quality guarantee for your documents.
- Preserve permissions. Similarity ranking is not authorization. Filter results by the requesting user’s access and check removal of restricted or deleted material before exposing the search to staff.
The limits and cost are part of the decision
The Workers AI model documentation specifies an 8,192-token input window shared across the content in an embedding. It requires media data rather than remote media URLs, and documents task prefixes for text retrieval. The listed input price is $0.0017 per million tokens as checked for this article. That is not the complete cost of ingestion, vector storage or the application around it.
For a pilot, compare useful results at a fixed shortlist length and record failures, latency and total operating cost. Decide whether the shorter vectors are worthwhile from that evidence, not just their smaller storage footprint.
Limitations: This is an explanation of the announced availability and documented interface. We have not run the model, benchmarked retrieval quality or tested Hebrew accuracy. Using the hosted service also means reviewing what content your organization permits sending to that provider.