Cloudflare has introduced Clef and Clef-flash, two open-source AI decision models designed to help AI agents make faster, structured decisions without generating lengthy text responses.
Released Oct. 1, the models target enterprise tasks such as routing support requests, classifying websites, and deciding when an automated workflow needs human review.
Both models are available through Cloudflare Workers AI, with downloadable weights under an Apache 2.0 license for local deployment. Cloudflare also announced a reinforcement-learning fine-tuning service for organizations that want to adapt Clef to their own data and workflows.
Unlike conventional chatbots, Clef produces predefined answers with probability scores. A support system, for example, could classify a customer request by urgency and department, then use those results to determine where the ticket goes.
How Cloudflare's Clef decision models work
According to Cloudflare's launch announcement, Clef uses a different approach from AI models that generate responses word by word. It processes an input and evaluates valid choices in parallel, returning structured classifications rather than free-form explanations.
The larger Clef model uses a frozen Qwen3.8-27B backbone, while Clef-flash uses Qwen3.5-9B. Both employ specialized routing components and adapters designed to improve decision accuracy and speed.
Cloudflare says the models can classify domains, route customer-support requests, evaluate trust and safety submissions, and help distinguish legitimate web crawlers from potentially harmful bots.
The company claims Clef currently leads the Jev Decision Index and outperformed Typesafe AI's Jev decision model in three of four evaluated categories. Cloudflare also reports lower latency across most of its internal comparisons.
Those are company-reported results, rather than independently reproduced performance measurements.
Cloudflare's approach could be useful for enterprises running numerous automated workflows where classification speed and consistent output formats affect processing costs.
The release also introduces customization options. Cloudflare's reinforcement-learning service uses AI Gateway to collect workflow data, Workers AI to generate training runs, Containers to evaluate results, and a new Trainer component to update model weights.
Fine-tuned models can then be redeployed through Workers AI. Initial customer support comes through Cloudflare's engineering team, with a self-service fine-tuning platform planned for later.
Organizations considering local deployment can use the Apache 2.0-licensed model weights without adopting the full Cloudflare hosting stack. Customers seeking Cloudflare-managed tuning would use more of its AI platform.
What eWeek found: Faster decisions still need security controls
Clef addresses a specific part of an AI workflow: deciding how to classify information or which predefined option to select. It does not independently determine whether the resulting action is authorized.
For example, a model might correctly classify a customer request as urgent and recommend escalation. The surrounding application still needs to decide whether the agent has permission to access customer records, change account settings, or contact external systems.
In its separate Agent Access Model proposal, Cloudflare argues that agent actions should be authorized against the task and its current state, using restricted permissions rather than trusting an agent's decisions alone.
The distinction also applies to other enterprise AI systems. eWeek has previously covered how Dataiku tracks AI agents across platforms and how Nvidia's agent safety system aims to restrict autonomous activity.
Organizations evaluating Clef should therefore examine classification accuracy, response times, probability calibration, and the policies controlling what happens after a decision.
Probability scores can help applications route uncertain cases to human reviewers, but a high score is not an authorization to perform a sensitive action.
Independent deployment results will also help establish how Clef performs beyond Cloudflare's tests. Existing enterprise AI agent benchmarks have demonstrated substantial differences between model capabilities and successful completion of real workplace tasks.
For lower-risk activities such as support-ticket routing, faster structured classifications could reduce processing overhead. Security-sensitive decisions involving fraud, customer data, or access controls require additional safeguards.
Clef gives developers another way to handle classification within AI agent workflows. Whether enterprises adopt it broadly will depend on how accurately its probability scores perform on their own data and how reliably downstream systems enforce authorization and approval requirements.
Want to learn more AI tips, tricks, and prompting techniques? eWeek readers get free 7-day access to The Neuron Academy, our practical learning platform designed to help professionals use AI more confidently at work. Browse all lessons →
Let us teach you How to Talk to AI for free! Try our six-minute course at The Neuron Academy and learn a few simple ways to write better prompts and get more useful results from AI, or browse our other AI course for free for seven days. Check out all the lessons here →