Cloudflare launched two new decision models named Clef and Clef-flash on Thursday. The release comes two weeks after TypeSafe's Jev model launched. Cloudflare claims its new models are smarter and faster than Jev. Both Clef models work similarly to Jev by answering bounded, structured questions in yes or no, multiple choice, and ranking formats.
Clef uses a large language model backbone for its processing. Cloudflare stated that Clef uses specially post-trained, frozen versions of Qwen3.8-27B and Qwen3.5-9B for Clef and Clef-flash. The Qwen backbone performs a prefill-only pass during inference. Clef then scores choices in parallel. TypeSafe keeps the underlying architecture of Jev secret.
Cloudflare claims Clef is more accurate than Jev
Cloudflare tested Clef against Jev and other open decision models using the Jev Decision Index on Hugging Face. The company claims Clef is slightly slower than other open models but more accurate. Clef-flash matches the accuracy of most competitors while offering higher speed. Cloudflare self-reported these scores, and they await independent reproduction on the official Decision Index.
Cloudflare also claims Clef beat Jev in three out of four areas on TypeSafe's own benchmarks. Jev won only in agent trace observability. Clef handles images and video in addition to text. It also supports a 64k context window. Jev handles up to 64k tokens per request, but its state and longest individual question limit is 32k.
Clef costs nearly six times more than Jev
Developers can access Clef directly through Cloudflare hosted on Workers AI. Cloudflare states edge GPUs reduce network latency and speed up decisions. Clef costs $0.24 per million tokens. That price is nearly six times higher than Jev at $0.042 per million tokens.
Users can download Clef from Hugging Face under the Apache-2.0 terms used by Qwen. Cloudflare AI Platform group product manager Michelle Chen confirmed the training datasets are not public. Clef-flash requires a GPU with at least 41 GB of VRAM. The larger Clef model requires 85 GB of VRAM. These requirements assume single concurrency and a 64k context window.
The API for Clef is fully Jev compatible. Developers can use it as a drop-in replacement locally or through Cloudflare hosting. Cloudflare plans to offer the models for local execution and managed cloud deployment.
