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

Cloudflare released Clef and Clef-flash, open-weight decision models hosted on Workers AI. Unlike standard LLMs, these models provide bounded, typed, probability-scored outputs instead of open-ended text, enabling direct workflow automation for tasks like support ticket routing. Clef (27B) and Clef-flash (9B) are Jev-API compatible and support image inputs. They deliver median latencies of 209.3 ms and 38.8 ms, respectively. Cloudflare also introduced a new reinforcement learning fine-tuning product to accompany these specialized decision models.