Original caption
Cloudflare just released Clef, the first models its Workers AI team trained itself, and they're open source. Clef and Clef-flash are decision models. You send some input and a few questions, and they return typed answers with probabilities instead of text: yes or no, one choice from a list, or a score. Your code uses that to route a ticket, escalate, or hand off to a person. Clef is 27B, built on Qwen3.8-27B. Clef-flash is 9B, built on Qwen3.5-9B. Both take 64K tokens of context and up to 4 images, and Jev only reads text. In Cloudflare's own tests they lead 7 of 10 decision benchmarks, like BANKING77 at 94.2 against Jev's 79.7. Jev still wins When2Call and BRIGHT, and keeps agent trace observability in Typesafe's own suite. Speed: Clef-flash has a median of 38.8 ms, Clef 209.3 ms, Jev 524.1 ms. Laya is faster still at 5.8 ms, but scores far lower. It's on Workers AI at $0.24 per million input tokens for Clef and $0.09 for Clef-flash, fully Jev-API compatible, and the weights are on Hugging Face under Apache 2.0. Cloudflare is also starting an RL fine-tuning service, hands-on first and self-serve later. Every benchmark here is Cloudflare's own, so I'd test it on your data before swapping. Source: blog.cloudflare.com/clef-decision-models Independent briefing by CodenamePoshan, not a Cloudflare post. #Cloudflare #AIAgents #OpenSource #MachineLearning #AIEngineering #DeveloperTools