Cloudflare has released Clef and Clef-flash, two decision models designed to return structured answers with probabilities rather than open-ended prose. For developers building AI agents, the useful distinction is between deciding where a task should go and generating whatever comes next. The models are hosted on Workers AI, with downloadable weights under an Apache 2.0 licence. That gives developers both a managed service and a route to experimenting locally.
Cloudflare Watch analysis
What happened
In its 1 October announcement, Cloudflare introduced the two models as compatible with the Jev API and also announced reinforcement-learning fine-tuning to adapt Clef to particular use cases.
A decision model handles a bounded question. Given a customer-support message, for example, it could classify whether the request is urgent and which team should receive it. The output supplies typed answers and probabilities that application code can use to route the ticket, trigger an escalation or refer it to a person.
Cloudflare describes these models as able to work with different classification categories without retraining for each new set. Its own testing includes using Clef to classify website domains, illustrating a job where software needs categories rather than a paragraph explaining its feelings about the internet.
Read Cloudflare’s Clef announcement.
Why it matters
An agent workflow contains plenty of decisions that do not require a full written response: choose a queue, assign a category, flag an exception. A model built around structured answers gives developers a different tool for those steps, while leaving generative models to handle writing and other open-ended work.
API compatibility makes comparison with Jev easier, while the downloadable weights let teams explore running the models themselves. Neither removes the need to test how well the classifications fit the actual job.
Our read
This is a worthwhile addition to the agent toolkit because it gives a model a narrower job description. Not every fork in the road needs a novelist.
Start with one decision your application already makes, such as ticket routing. Compare the model’s classifications with your existing process, measure response time and cost, and decide which results require a person’s attention. A probability is an input to that policy, not permission to dispense with it.
The fine-tuning announcement is also worth following: adapting a decision model to a specific workflow could be more useful than asking a general-purpose assistant to become an expert through an increasingly elaborate prompt.
What to watch
- How Clef and Clef-flash compare on real classification tasks, including incorrect decisions and referrals to people.
- What access, pricing and training requirements Cloudflare sets for reinforcement-learning fine-tuning.
- Whether developers find the hosted service or locally run weights the better fit for their workloads.
Discussion spark: Should routine agent decisions move to specialised classification models, or is one general-purpose model easier to test and maintain?
Sources and evidence
- Source update (1 October 2026, 15:34 UTC)
not affiliated with or endorsed by Cloudflare