Cloudflare Clef: Open Decision Models for AI Workflows

Cloudflare released Clef and Clef-flash on Workers AI on October 1, 2026. They are decision models: instead of drafting prose, they take an input state and questions with predefined answer types, then return structured answers and probabilities. For a business, the immediate opportunity is narrow workflow triage—such as classifying an incoming request and deciding whether it needs human review—not replacing the judgment behind that workflow.

The release matters because Cloudflare now offers hosted decision models and downloadable model weights. It gives teams another option to evaluate alongside TypeSafe AI's Jev. Cloudflare says Clef follows the System One API used by Jev, but compatibility and output quality should be tested against a real integration before any switch.

What did Cloudflare release?

The two hosted models are @cf/cloudflare/clef and @cf/cloudflare/clef-flash. Cloudflare positions Clef as the higher-precision option and Clef-flash as the faster option for latency-sensitive paths. Its documentation lists 27-billion and 9-billion parameter sizes, respectively, and a 64,000-token context window for each. Both can be called through Workers AI. Cloudflare also published their weights on Hugging Face.

The models answer up to 64 typed questions in a request. The supported question types are a yes/no probability (noul), a choice among developer-defined options (choice), and a score against an ordered rubric (score). An application can use those outputs in its own rules. The model does not itself define what action is appropriate for the business.

Cloudflare separately announced a reinforcement-learning fine-tuning service. Its launch materials invite customers with relevant use cases to join as design partners. That should not be read as a generally available self-service fine-tuning product.

Where could a business use a decision model?

Consider a shared inbox that receives sales questions, support requests and incomplete submissions. A team could define a small set of categories, ask whether required information is present, and route only high-confidence matches automatically. Uncertain or consequential requests would go to a person. That is a possible pilot design, not evidence that Clef is accurate enough for a particular organization.

Other reasonable starting points include classifying internal requests, tagging documents for review, or deciding which existing workflow branch should handle a routine item. In each case, the business should own the categories, exception handling and approval rules. A probability is a useful signal, not permission to make an irreversible decision.

Is Clef faster or more accurate than Jev?

Cloudflare publishes benchmark results comparing Clef, Clef-flash and Jev, including latency and decision-task scores. Those are Cloudflare's own evaluations, not an independent guarantee of performance on your data. Model rankings can change with the questions, inputs, deployment path and failure costs of a specific workflow.

For a useful comparison, create a representative test set from permitted business data. Measure decision accuracy, probability calibration, latency, cost, and the number of cases that still need manual review. Include ambiguous and out-of-scope inputs. Compare the whole workflow, not just an isolated model response. The most important result is whether the new approach improves a measurable operation without creating unacceptable mistakes.

What should a pilot check before going live?

1. Define the decision boundary. Write down the allowed answers, the action tied to each answer, and which cases always require a human.
2. Test representative examples. Include routine cases, edge cases, missing context and intentionally ambiguous requests.
3. Validate the integration. If migrating from Jev, check the actual request and response contracts, error handling and thresholds; do not rely on a compatibility claim alone.
4. Keep a safe fallback. Log model outputs, monitor exception rates and make it easy to route uncertain work to a person.
5. Review data handling. Confirm what information the workflow sends to the selected hosting environment and apply the organization's data policies.

The practical takeaway is that decision models are becoming a more concrete tool for bounded automation. Clef adds a Cloudflare-hosted and downloadable option, but a successful deployment still depends on good workflow design and evaluation.

Oplix helps teams design AI-enabled workflows, automate repeatable operations and build the software around them. If you have a specific classification or routing task, talk with us about a small, measurable pilot before putting a model in the decision path.