Eigen RadarAI
Archive
Analysis

Cloudflare’s Clef models score choices instead of writing an answer

Cloudflare released Clef and Clef-flash on Workers AI and published their weights under Apache 2.0. The decision models score choices supplied by a developer rather than generating a free-form answer, supporting tasks such as routing requests and selecting tools. Images can accompany the input. Initial customer fine-tuning engagements involve Cloudflare engineers.

Artificial Intelligence··Midday
A computer screen shows three distinct natural texture images with separate confidence bars.

Clef scores the answers an application permits

Cloudflare released Clef and Clef-flash on 1 October through Workers AI, its hosted AI service, and published their weights under Apache 2.0. These decision models read a supplied state and a schema of questions, then return probabilities for the permitted answers. Their output can support classification, request routing and tool selection inside an application.[1], [2]

A vision encoder lets developers include images alongside text. The models preserve compatibility with the Jev application programming interface, allowing applications built for that decision-model format to use the same request structure. Choices are scored in parallel while the input is processed, without intervening free-form text generation.[1]

Training adapts the decision layer

Clef uses a Qwen3.8-27B backbone, while Clef-flash uses Qwen3.5-9B. Those backbone weights remain frozen as adapters and a decision-routing head are trained. Training examples vary the order of fields, prompt formatting and answer schemas. Attention between fields lets the system handle several decisions within one schema.[1]

Customer tuning begins with Cloudflare engineers

Cloudflare also introduced an adaptation workflow in which AI Gateway captures requests, Workers AI runs trials and Containers scores outputs. A Trainer component updates the model using those scores before deployment. Initial customer engagements involve Cloudflare engineers. The company says hosted customer requests are excluded from training unless the customer explicitly opts into fine-tuning.[1]

References

  1. News sourceCloudflare BlogCloudflare releases Clef models for structured decisions↩1↩2↩3↩4
  2. News sourceMarkTechPostCloudflare releases Clef and Clef-flash for decisions among specified answers↩
AI safety and regulationChip wars and the supply chain

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →