Cloudflare releases Clef, open-weight “decision models” that answer agents with a fixed choice instead of free text

On 1 October Cloudflare released two “decision models”, Clef and Clef-flash. Instead of writing free-form answers, they score a fixed set of allowed answers, such as yes or no, multiple choice or a ranking, and return a probability for each. Cloudflare pitches them for the points in an AI agent's workflow where a decision is needed: routing a support ticket, triggering an escalation or handing a case to a human. They are built to work with the same API as Jev, a decision model from Typesafe AI.

Both are built on frozen Qwen backbones that Cloudflare post-trained with low-rank adapters. They run on Cloudflare's Workers AI service, and the weights are on Hugging Face under the Apache 2.0 licence. In Cloudflare's own benchmark, median latency was 209.3 ms for Clef and 38.8 ms for Clef-flash, against 524.1 ms for Jev. “We also ran benchmarks across Typesafe's own eval suite and our Clef models fared well, beating Jev in 3 out of 4 areas,” Cloudflare wrote. It also opened a reinforcement-learning fine-tuning service so customers can tune Clef on their own data.

The Register reported that Cloudflare calls the models open source but has not published the training data, so they are open-weight, and that the scores have not yet been reproduced on the independent Decision Index. It also reported a price of $0.24 per million tokens, against $0.042 for Jev; Cloudflare's post gives no price.

• Confirmed Clef and Clef-flash were released on 1 October as open weights under Apache 2.0, on Hugging Face and Workers AI. Cloudflare
• Claimed Median latency is 209.3 ms for Clef and 38.8 ms for Clef-flash, against 524.1 ms for Jev, in Cloudflare's own benchmark. Cloudflare
• Reported The training data is not public and the scores have not been reproduced on the Decision Index, The Register reported. The Register