Clef

Clef is a 27B multimodal decision model developed by Cloudflare and available on Workers AI as @cf/cloudflare/clef. Rather than generating free-form text, it reads a state and a schema of typed questions and returns a probability for every allowed option. The Cloudflare changelog dated October 1, 2026 describes it as the first model trained by the Cloudflare Workers AI team, with weights published on Hugging Face.

What It Is

Clef is a decision model in the same family as Typesafe's Jev, according to the changelog. It accepts a state as text, JSON, images, or video, plus up to 64 typed questions per request. The model page lists a 65,536-token context window and vision support. Output is structured, so an application can act on the result directly, for example by routing a ticket, blocking a request, or escalating to a human, without parsing free-form output.

How the Question Types Work

The changelog documents three question types:

• noul: a yes/no question that returns the probability of yes.
• choice: pick one option from a set you define, returning the chosen option, per-option probabilities, and a confidence value.
• score: rate against an ordered rubric, returning a probability-weighted score and a probability per level.

Images can be supplied with the state (up to four PNG, JPEG, or WebP files). Remote image URLs are not accepted.

Speed and Benchmarks

Cloudflare states that Clef is built for the hot path. In its 43 benchmark runs, the changelog reports median latency of 209.3 ms for Clef versus 524.1 ms for Jev, and says Clef is 2.5x faster at the median. Cloudflare also reports that across 10 decision benchmarks a Clef model scores highest on 7, and that Clef beats Jev in 3 of 4 areas on Typesafe's workflow evals. These are vendor-published results; the full figures are on the Hugging Face model card.

Use Cases and Compatibility

Cloudflare lists support triage, threat intelligence, trust and safety scoring, agent guardrails, and visual classification as example uses. Clef follows the System One API, so the changelog says an existing Jev integration can be switched by changing the endpoint and model. A smaller 9B sibling, clef-flash, targets latency-critical decisions.

Access and Setup Path

Clef is called through the Workers AI binding with env.AI.run() or the REST API at /ai/run, and AI Gateway can be used with these endpoints. Cloudflare says the weights are open-sourced under the Apache 2.0 license on Hugging Face, and that it is launching a reinforcement learning fine-tuning service with design partners.