Workers AI adds EmbeddingGemma 2: one embedding model for text, images, video and audio (November 12, 2026)
@cf/google/embeddinggemma-2 puts text, images, video and audio in one 768-dimension space for $0.0017 per 1M input tokens.
What changed
On November 12, 2026 Cloudflare added @cf/google/embeddinggemma-2 to Workers AI. It is an open embedding model from Google DeepMind that maps text (including code), images, video and audio into the same 768-dimension vector space. One text query can then search a mixed collection of documents and media.
- Price: $0.0017 per million input tokens.
- Context: 8,192 tokens, shared by everything in one input. Cloudflare's model page says an image uses about 260 tokens, video about 140 tokens per sampled frame (one frame per second) and audio about 25 tokens per second.
- Shorter vectors: set
dimensionsto 128, 256 or 512 (the default is 768). Queries and documents must use the same size. - Media goes in as base64 data URIs. Remote URLs are not supported.
- Task prefixes improve results. For search, use
task: search result | query: ...for queries andtitle: ... | text: ...for documents.
Prerequisites
- A Cloudflare Worker with the Workers AI binding (
AI). No API key is needed inside the Worker.
Steps
Text, from the model page:
export interface Env {
AI: Ai;
}
export default {
async fetch(request, env): Promise<Response> {
// Prefix queries and documents with task instructions for better retrieval quality.
const response = await env.AI.run("@cf/google/embeddinggemma-2", {
input: [
"task: search result | query: waterproof shoes for trail running",
"title: Trail Runner X | text: Waterproof running shoes with a breathable mesh upper.",
"title: City Sneaker | text: Lightweight canvas sneakers for everyday wear.",
],
});
// response.data[i].embedding is a 768-dimensional vector
return Response.json(response);
},
} satisfies ExportedHandler<Env>;