Cloudflare Tries to Outplay Jev With Open-Weight Clef Models

Cloudflare has introduced two open-weight decision models, named Clef and Clef-flash, which are capable of handling structured yes/no, multiple-choice, ranking tasks, as well as image and video processing, with a context window of up to 64K. According to Cloudflare, these models outperform TypeSafe's Jev on several benchmarks and can run locally from Hugging Face.

Clef is built on an LLM backbone, utilizing specially post-trained, frozen versions of Qwen3.8-27B for Clef and Qwen3.5-9B for Clef-flash. The Qwen backbone performs a prefill-only pass during inference. Despite not being as fast as Jev, Clef manages to score choices in parallel after the prefill-only pass, making it faster. The company's own ranking suggests that Clef is slightly slower than other open models but more accurate, while Clef-flash matches the accuracy of most models, yet is far faster.

Cloudflare ran Clef against Jev and other open decision models using the Jev Decision Index, which is available on Hugging Face. Although the exact benchmarking of Clef's scores is yet to be reproduced, the company claims that Clef has outperformed Jev in three out of four areas, with the only loss being in agent trace observability.

Although Clef might be a bit slower or less accurate, it has a significant advantage over Jev - its ability to handle images and video, and its support for a larger 64K context window. Jev can also process up to 64k tokens across a request, but its state and longest individual question are limited to 32K.