Cloudflare's Clef is a small model built to decide, not to chat

Agent workflows spend a lot of calls on small decisions: which branch to take, which category a message fits, whether a document is relevant. Using a full chat model for those is slow and expensive. Cloudflare's answer is Clef, an open-weight decision model that does not write free text at all. It takes a schema of choices and returns a calibrated probability for each, which makes its output bounded and easy to use in code.

The design is where it gets interesting. Clef is built on Qwen backbones, 3.8 billion parameters for the main model and 9 billion for the faster Clef-flash, and it runs a non-autoregressive, prefill-only pass that scores every option in parallel rather than generating tokens one at a time. That shows up in latency: Cloudflare reports a median of 209 milliseconds for Clef and 38.8 milliseconds for Clef-flash, against 524 milliseconds for the earlier Jev model, and it beats Jev on three of four workflow tests. It carries a 64K context window and a vision encoder for image classification. Training leans on label-smoothed cross-entropy and a calibration method the company calls RLCD, with synthetic data that permutes field orders so the model does not memorize schema layouts.

Alongside the model, Cloudflare is opening a reinforcement-learning fine-tuning service that starts with hands-on help from its engineers and moves toward self-serve training on its existing AI Gateway and Workers AI infrastructure.

Why it matters

If you build agents, the routing and classification steps are often the hidden cost, and a cheap, calibrated model that answers in tens of milliseconds lets you reserve the expensive chat models for the work that actually needs open-ended reasoning.