Cloudflare open-sources Clef decision models, 13x faster than Jev
Apache-2.0 models that return confidence-scored choices instead of text, plus an RL trainer. Amazon shipped its own open 2B decider the same week.
Copy markdown
Clef and Clef-flash land on Workers AI, Apache-2.0
Clef (Qwen-27B base, rank-256 adapters, 64k context, plus a vision encoder) and the latency-tuned Clef-flash (Qwen-9B) return a typed choice with a calibrated confidence score instead of generating text. Clef-flash answers in a 38.8ms median versus Jev's 524.1ms, and Clef tops the Jev Decision Index. You can swap a brittle “LLM-as-router” prompt for a deterministic classifier on support triage, bot detection, or trust-and-safety calls — run it on Workers AI or pull the weights from Hugging Face.
A new RL trainer tunes the decider on your own traffic
Clef ships with a reinforcement-learning platform that captures real agent traffic through AI Gateway, generates rollouts on Workers AI, scores them in Containers, retrains weights with a new Trainer, and redeploys your custom model — all inside Cloudflare. It's white-glove with their engineers for now, with self-serve promised later, but it closes the loop from “log your decisions” to “a model tuned to them” without you standing up a GPU pipeline.
Amazon's Strands Decider 2B is small enough to self-host
AWS's Strands Labs open-sourced its own decision model the same week — a 2B model built on Qwen3.5-2B that briefly topped Jevbench for its size. At 2B it runs locally with low latency and no cloud dependency, giving agent builders a free, embeddable routing layer rather than a hosted API call. Grab the weights from the Strands Labs GitHub.
Why decision models are suddenly everywhere
The category TypeSafe's Jev kicked off — a hallucination-free model that returns one of your schema's options with a probability, not a paragraph — is becoming a commodity primitive. Instead of begging a chat model to “pick one and don't explain,” you call a tiny, auditable model that's cheaper and faster. With Cloudflare and Amazon both open-sourcing weights in days, a calibrated agent router is now a weekend drop-in.