A decision model from the company that routes a fifth of the web. Cloudflare shipped Clef (27B) and Clef-flash (9B) on October 1, 2026, Apache 2.0, fine-tuned from the Qwen3.5 27B backbone. Both are decision models in the Jev sense: hand one a question with a fixed set of answers (which inbox does this email route to, is this request abusive, which of these five actions should the agent take) and it returns one answer with a confidence number instead of a paragraph. Clef adds an image input, so a screenshot of a dashboard is a legal question. In a week Clef collected 1,859 likes and Clef-flash 17,600 downloads on Hugging Face.
What the flash size buys. Clef-flash at 9B is the laptop-runnable end of the decision-model map: about 5.4GB in 4-bit, which is a comfortable fit on a 16GB machine with the OS still loaded. Clef at 27B wants more: about 15.7GB in 4-bit, still one-card territory but with a KV cache that matters at length - 256MB per 1,000 tokens means a full 262k context costs about 67GB of cache on top, so the honest serving statement is that the 27B card runs whole-document decisions at 32k context in about 24GB, and short-prompt routing anywhere Qwen3.5 27B already runs.
Why a CDN company ships models. Cloudflare runs Workers AI as a serving platform and the decision-model shape (fast, typed, cheap at scale) is the exact shape a router wants for abuse detection, request classification, and agent-tool gating at the edge. The open weights are the credibility move: any customer can read what the classifier does before trusting it with their traffic, which is the same trust argument Clef’s own blog makes.