Decision models answer a fixed set of questions about an input instead of writing sentences, and Cloudflare has now built two of its own. Clef and Clef-flash are the first models to come out of the company’s Workers AI team, and both are open-weight under Apache 2.0.
Neither generates free-form text. A caller supplies a state and typed questions, and the model returns a probability for each permitted answer. The pair speaks the same API as Jev, the TypeSafe AI model that opened the category in September, so switching amounts to changing an endpoint and a model name. Clef runs on Workers AI, with weights on Hugging Face for self-hosting. Requests can carry up to 64 questions and four images.
Both models began life elsewhere. Qwen3.8-27B supplied the backbone for Clef, and the slimmer Qwen3.5-9B did the same for Clef-flash; each kept its source model’s vision encoder through post-training. Inference then splits in two. The backbone makes a single prefill pass, and a joint schema head reads the final states, routes evidence to each question, lets fields attend to one another, and scores every option at once.
Results are mixed by design. On Cloudflare’s ten-benchmark shortlist a Clef model topped seven, yet Jev stays far ahead on difficult reasoning, scoring 78.3 to Clef’s 48.0 on GPQA Diamond. Speed is Cloudflare’s pitch: in a threat-intelligence test Clef judged a domain in 2.2 seconds against 4.7 for gpt-oss-120b.
