Cloudflare has released Clef and Clef-flash, AI models that score a set of permitted answers rather than compose a chat response. Both became available on October 1, 2026, through Workers AI, Cloudflare’s hosted model service, and as downloadable releases.
A support application, for example, could supply a customer’s message and ask whether billing, technical support or sales should handle it. Clef returns probabilities for those choices, which software can use directly. This gives an AI agent a classification step before it calls a tool or passes work to another team.
How Clef turns context into a bounded answer
The developer defines the questions and the answers each one permits. Supported formats include yes-or-no questions, a choice among named categories and ratings on an ordered scale. The hosted Workers AI interface accepts up to 64 questions in one request and up to four attached images, allowing visual content to be classified too.
The model card explains that a Qwen backbone processes the supplied context. An added scoring component uses that information to evaluate the allowed options together, skipping the step of writing an explanation one text fragment at a time.
Cloudflare reports internal speed and classification tests against other models. Their results apply to the workloads and conditions tested. They do not establish the latency or error rate a customer will get from a different collection of messages or images.
A probability still needs an application policy
A high probability is the model’s estimate. Assessing whether it is dependable for a particular task requires testing cases with known correct answers. Cloudflare describes training intended to improve probability calibration, meaning how closely its estimates match observed results, without establishing reliability for every customer’s use case.
Application code can use the scores to route a ticket, or ask a person to review an uncertain answer. The application builder sets that rule. The hosted API documentation does not specify a universal confidence cutoff or a human-approval mechanism built into the model.
Hosted prices and downloadable releases
The Clef and Clef-flash downloads are released under Apache 2.0, allowing developers to run and adapt them outside Workers AI. Running a download means setting up the software and providing hardware that can load the model. It is not a ready-made consumer app for any laptop.
Workers AI lists input pricing of $0.24 per million tokens for Clef and $0.09 for Clef-flash. Tokens are the units used to account for the content a model processes. These are model-service rates, rather than the full cost of operating an application.
Cloudflare is also offering hands-on engineering help to tune Clef for individual business workloads. A self-service platform for customers to train and redeploy their own versions is planned for later.
The conversation starts here
Sign in with a supporter account to comment. Sign in
Nobody has commented yet. Want to go first?