Clef-flash vs Jev: format matches, speed claim fails

Clef is the first model Cloudflare's Workers AI team trained itself, built on Qwen3.8-27B and Qwen3.5-9B backbones and released alongside a reinforcement-learning fine-tuning pipeline. That pipeline stitches together AI Gateway, Workers AI, Containers and a new Trainer product, but only a design-partner program with Cloudflare's field engineers is open today, with no self-service timing or pricing given.

The benchmark picture is mixed rather than a clean win. On the community Jev Decision Index, Cloudflare reports Clef in first place, but the leaderboard carries a "Self-reported" note and the scores have not yet been reproduced on the original community board. Clef leads on classification and intent tasks such as BANKING77 and API-Bank, while Jev is far ahead on knowledge and multi-step reasoning such as GPQA Diamond and BBH. Clef-flash is fast but loses considerable accuracy on out-of-scope detection and hallucination detection, so it is not simply a faster Clef.

The clearest differentiators appear to be image input, which Jev does not support, and the openly licensed weights that let a team run both models locally with the same answers as the cloud. What the test ultimately suggests is that the gap between these models is smaller than the gap made by how well instructions and criteria are written, so the choice is likely to hinge less on the model and more on whether a team has already documented its own decision rules and threshold behavior.