Cloudflare Clef and Clef-flash: decision models for agent hot paths in interviews

On October 1, 2026, Cloudflare announced Clef and Clef-flash, the first models trained by the Cloudflare Workers AI team. Both are hosted on Workers AI as @cf/cloudflare/clef and @cf/cloudflare/clef-flash. For interview prep, the useful story is practical: when an agent needs a fast, typed, calibrated choice on a hot path, a decision model can beat waiting on free-form generation.

A decision model is not a chat LLM with a stricter prompt. You pass an input state plus typed questions. The model returns probabilities over allowed answers. There is no free-form text generation and no long chain of reasoning tokens to wait for. That matters when the job is classify, route, approve, or score against a fixed schema rather than write an explanation.

Cloudflare positions the family against Typesafe’s Jev and the System One API. Clef is fully Jev and System One compatible, so teams can try a drop-in by changing endpoint and model. Clef is about 27B parameters, post-trained from Qwen/Qwen3.8-27B, and aimed at highest-precision decisions. Clef-flash is about 9B, post-trained from Qwen/Qwen3.5-9B, and aimed at latency-critical hot paths. Both are multimodal with a vision encoder. On Workers AI, state can be text, JSON, images, or video. Optional images are capped at four (PNG, JPEG, or WebP