Cloudflare releases Clef, a 27B schema-driven decision model

Cloudflare has released Clef, a 27-billion-parameter open-weight model that returns probabilities over typed answer options rather than free-form text. According to the Hugging Face model card, Clef "turns a state and a schema of typed questions into decisions" and "reads the state as text, JSON, images, or video, and returns a probability for every allowed option of every question in a single forward pass."

The card is explicit on the point: "There is no free-form text generation and no output parsing." The backbone is Qwen3.8-27B with its vision encoder, topped with "a small transformer head that reads the backbone's final hidden states, routes evidence from the state to each question, and scores all options of all questions jointly." Three question types are supported: noul (true/false), choice, and score. The weights ship under Apache-2.0.

Benchmarks on the card put MMLU accuracy at 90.3%, GSM8K at 80.8%, WinoGrande at 93.5%, and BFCL case-exact accuracy at 98.5%, with median latency of 209.3 ms and p95 of 238.6 ms. Cloudflare also publishes workflow numbers: 86.2% primary-action accuracy on invoice processing, 76.3% exact-actions on customer service, and 62.9% exact-actions on security incidents.

"The Clef API is fully compatible with Jev and SystemOne," the card notes. A pair of researchers in our tracker had already posted the page by the time this went out.