Cloudflare released two “decision models,” Clef and Clef-flash, on 2026-10-01, hosting them on Workers AI and “fully open-sourcing these models on Hugging Face under an Apache 2.0 license.” A decision model, in Cloudflare’s framing, produces “bounded structured outputs cheaply, quickly and consistently” for the points in an agent workflow where a choice is needed, as opposed to an LLM that generates open-ended text. Cloudflare credits the recent buzz around Typesafe AI’s Jev model for the category, says Clef is “fully Jev-API compatible,” and claims Clef currently leads the Jev Decision Index. It also announced a reinforcement-learning fine-tuning product, offered first through its forward-deployed engineers and later as self-serve.
The models are built on frozen Qwen backbones: Qwen3.8-27B for Clef and Qwen3.5-9B for Clef-flash, with a trained routing head and rank-256 low-rank adapters. At inference Clef runs a prefill-only pass and then scores the valid schema choices in parallel; “The decision step is non-autoregressive,” so no intermediate text is generated and outputs are probabilities over allowed options. Training used label-smoothed cross-entropy with a Brier loss for calibration, plus a method Cloudflare calls Reinforcement Learning for Calibrated Decisions (RLCD). Unlike Jev, Clef has a vision encoder and a 64k context window (Jev: 32k).
In Cloudflare’s own threat-intelligence use, Clef classified a website domain into categories with probabilities in 2.2 seconds including fetch and render, against 4.7 seconds for gpt-oss-120b in the same workflow, which also returned fewer classifications.
Why it matters: it is a notable open-weight entry in a fast-forming category, small calibrated models that make typed decisions inside agent pipelines, and a clear case of a Chinese open-weight base model being productized by a U.S. infrastructure company. What it does not show: the benchmark comparisons and the latency example are Cloudflare’s own, the Jev Decision Index is a young third-party benchmark, and the post does not report how the models behave on decision schemas far from its training distribution.