Cloudflare released the Clef and Clef-flash decision models on Workers AI on October 1, 2026. Their weights are also available on Hugging Face under the Apache 2.0 license. Unlike a general large language model that generates open-ended text or tool calls, Clef is designed to return predefined, typed choices with probabilities. Source

Use it for classification and routing

A team can submit a support message and ask whether it is urgent, which team should handle it and how severe it is. Code can then route the request, escalate it or defer to a person when confidence is insufficient. This makes a decision model relevant when a workflow repeatedly classifies inputs within known choices, rather than producing new explanatory text. Source

The models accept images as well as text, have a 64,000-token context window and are compatible with the Jev API. Tokens are the units used to process input; the context window limits how much information a decision can consider together. Cloudflare published its own accuracy and latency benchmarks. Teams should still test representative data and consider the cost of incorrect classifications before putting a model into an automated path. Source

Choose a model and check availability

Cloudflare positions Clef for decisions that prioritize precision and Clef-flash for latency-sensitive work. Both can be called through Workers AI now, while the published weights allow local testing. Selection should account for response time, the impact of a wrong decision and the point at which a person reviews the result. Source

Fine-tuning adds training on examples for a particular workload. Cloudflare says it does not read, store or train on hosted requests or responses unless a customer uses fine-tuning. Workload-specific fine-tuning is currently offered as a hands-on service with Cloudflare engineers. A self-service platform for training and redeployment is planned later; the announcement does not specify its release date or price. Source