Workers AI

What it is

Workers AI runs machine-learning models on Cloudflare's own GPUs, available to your Worker through a single AI binding. You call env.AI.run("@cf/...model") with your input and get a result — no separate inference server, no API keys to a third party, no GPU to provision. The catalog spans text generation, embeddings, image generation, speech, and translation.

Strengths

• One binding — no external API keys or separate inference infrastructure.
• Runs on Cloudflare's GPU network, close to your Worker and your users.
• Broad model catalog: LLMs, embeddings, image, audio, and more.
• Pay per use; a free allocation makes prototyping cheap.
• Pairs naturally with Vectorize for retrieval-augmented generation.

Trade-offs

• Model selection is Cloudflare's catalog, not every model on the market.
• Largest frontier models may not be available; quality varies by model.
• Inference latency and rate limits apply, especially on bigger models.
• For a specific proprietary model, you'd still call that provider's API.

When to use it

Use Workers AI for in-app inference where you want it co-located with your edge logic: chat features, summarization, classification, embeddings for search, or image generation — without running your own GPU stack.

Vibe coding fit

Workers AI removes most of the setup an agent would otherwise script: no key management, no provider SDK, just a binding and a run call. Tell the agent which task and model you want (e.g. an embeddings model to feed Vectorize) so it picks from the catalog correctly. The example binds AI and runs a text model.

# wrangler.toml
[ai]
binding = "AI"

// inside your Worker
const out = await env.AI.run("@cf/meta/llama-3.1-8b-instruct", {
  prompt: "Summarize this in one sentence: ...",
});