Cloudflare Releases Clef and Clef-flash: Open-Weight Decision Models That Return Typed Probabilities Instead of Text

Cloudflare has shipped Clef and Clef-flash, its first models from the Workers AI team. They are not chatbots. Instead of generating a free-form reply, they take an input state plus a schema of typed questions and return probabilities for the allowed answers. That makes them feel less like a text generator and more like a machine for structured decisions.

The pair is open-weight under Apache 2.0, and both are compatible with TypeSafe AI’s Jev API. Switching from Jev is described as a simple endpoint-and-model-name change. Cloudflare says both models are live on Workers AI now, and the weights are also on Hugging Face for self-hosting.

Clef supports three question types: noul for yes-or-no answers, choice for selecting one named option with per-option probabilities and a confidence value, and score for rating against an ordered rubric. On Workers AI, a single request can include up to 64 questions and up to 4 images. The models also use a 64,536-token context window, which is plenty of room for messy real-world inputs.

Under the hood, Clef is post-trained from Qwen3.8-27B and Clef-flash from Qwen3.5-9B. Cloudflare froze both backbones and trained a schema-routing head on top, using rank-256 low-rank adapters plus a mix of label-smoothed cross-entropy and Brier loss. It also added a secondary training objective called Reinforcement Learning for Calibrated Decisions, which gives partial credit for adjacent ordinal choices.

Cloudflare’s own benchmark table shows the split personality clearly. On a 10-benchmark shortlist from the Decision Index 0.2.1 suite, a Clef model led on 7. It did especially well on BANKING77 and CLINC150+OOS, while Jev kept the upper hand on knowledge-heavy tests like GPQA Diamond, MMLU-Pro and BBH. In TypeSafe’s own workflow evaluations, Clef edged Jev in 3 of 4 areas, but only by small margins. None of the numbers have independent replication yet.