Two weeks after TypeSafe's Jev model appeared, Cloudflare has released its own pair of open-weight decision models, Clef and Clef-flash. Both answer the same bounded question types as Jev—yes/no, multiple choice, and rankings—but use a different architecture: post-trained, frozen Qwen3.8-27B and Qwen3.5-9B backbones that perform a prefill-only pass before scoring choices in parallel.
Cloudflare's own tests on the Jev Decision Index suggest Clef is slightly slower than other open decision models but more accurate, while Clef-flash matches most rivals' accuracy and runs far faster. Against TypeSafe's benchmarks, Cloudflare claims Clef wins three of four areas, losing only on agent trace observability. These results have not been independently reproduced, and Cloudflare acknowledged that its training datasets are not public despite the "open source" label.
Clef also handles images and video, not just text, and supports a 64k context window. It costs $0.24 per million tokens—nearly six times Jev's $0.042—but can be downloaded from Hugging Face under Apache-2.0. For local use, Clef-flash requires at least 41 GB of VRAM and Clef requires 85 GB, assuming single concurrency and a 64k context. The API is Jev-compatible, so it can serve as a drop-in replacement.