Cloudflare has introduced Clef and Clef-flash, its first open-source, open-weight decision models available natively on Workers AI. Unlike traditional large language models, Clef evaluates input states against typed questions to return precise numerical probabilities rather than free-form text. By removing text generation and intermediate reasoning tokens, the architecture delivers ultra-low latencies well-suited for hot-path application tasks. Built on top of a frozen Qwen backbone, the models support text, JSON, and visual inputs across a broad 64K-token context window. Cloudflare is also launching a reinforcement learning fine-tuning service to help developers optimize these decision models for custom production workloads.

Full report : The new Apache 2.0 models bypass free-form text generation to deliver fast, typed probabilities for instant automated agent actions.