Cloudflare launches open-weight models for typed probability decisions
Cloudflare has released Clef and Clef-flash, its Workers AI team’s first models. They are designed for fixed decision tasks rather than conversation: given an input state and a schema of typed questions, they return probabilities for permitted answers instead of free-form text. Clef supports yes/no questions, named-option choices and ordered rubric scores. A Workers AI request can contain up to 64 questions and four images.
Clef has 27B parameters and Clef-flash has 9B. Both use Apache 2.0 licensing, accept images, retain their underlying vision encoders and are compatible with TypeSafe AI’s Jev API. Their weights are available on Hugging Face, while hosted versions run through Workers AI, the REST API and AI Gateway. Cloudflare reports median Workers AI latency of 209.3 ms for Clef and 38.8 ms for Clef-flash.
In a vendor-reported shortlist of 10 benchmarks, a Clef model ranked first on seven. However, Jev performed better on knowledge-oriented tests including GPQA Diamond, MMLU-Pro and BBH. Cloudflare also announced a reinforcement-learning service for tuning Clef on private data, initially involving its forward-deployed engineers. The company says the service will later offer self-service access.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.