Cloudflare Clef: Open Decision Models Explained

Cloudflare released Clef and Clef-flash on October 1, 2026, as decision models hosted on Workers AI, with weights released under Apache 2.0. Their role is to return bounded, structured choices for a workflow rather than produce an unrestricted conversational answer.

That can be useful for routing a support request, classifying a document or choosing a tool. The narrower output is easier to integrate into software. It does not mean the decision is automatically correct, fair or appropriate for a consequential action.

What a decision model returns

Cloudflare’s model documentation describes Clef as a multimodal model that reads a state and a schema of typed questions, then returns probabilities for the allowed options. Inputs can include text, JSON, images or video. The declared choices become part of the application design: the model cannot select a missing category simply because that category would better describe the case.

Consider a support workflow with billing, technical and urgent-service options. A request can involve more than one issue. The developer must decide whether the schema allows multiple classifications, whether urgency is a separate question and what happens when the available information is insufficient. Those decisions shape the behaviour as much as the model selection does.

Probabilities need testing before they drive actions

A high probability is a model output, not a guarantee. Test whether high-confidence answers are actually more reliable on the organization’s own cases. Use a held-out set with ambiguous requests, unusual language and adversarial instructions embedded in documents. Evaluate the exact schema that will be deployed.

For a Canadian service desk, include English, French and mixed-language requests. Measure mistakes by their consequences: routing a routine question to the wrong team costs time, while overlooking an urgent outage can have a much larger effect. A threshold for human review should reflect that difference rather than maximize a single accuracy score.

Where open weights help—and where work remains

Open weights make local experimentation possible, subject to the model’s requirements and licence. They do not provide the whole production workflow. Teams still need input controls, monitoring, version management and a record of which decision led to which action. Self-hosting and a hosted API also create different operational obligations.

Cloudflare’s launch describes its own benchmark results and use cases. Those should be treated as vendor evidence. The practical adoption test is whether a constrained model improves a specific workflow while preserving an abstention route and a human escalation path when the case falls outside the expected categories.

Sources and editorial method

We use the linked original statements and reports to distinguish documented facts from our analysis. Company announcements describe their authors’ plans and claims; they are not independent verification of future outcomes. This article is dated and will be revised if material evidence changes. Read our editorial standards and corrections process.

Cloudflare Clef launch ↗ — October 1, 2026; model releases, open-weight licensing and vendor evaluation.
Cloudflare Clef model documentation ↗ — Input types, typed questions and probability outputs.