Cloudflare's Clef Shows Most AI Features Need a Classifier, Not a Chatbot

Cloudflare released Clef and Clef-flash on 1 October, two open source decision models hosted on Workers AI. They do not write paragraphs. They make a classification, such as which route a request should take, and return it with a probability.

The numbers are the story. Clef has a median latency of 209.3 ms and Clef-flash comes in at 38.8 ms, with a p95 of 122.4 ms. Both are Apache 2.0 licensed, built on Qwen models, and score competitively across 43 benchmarks. They use non-autoregressive scoring rather than generating tokens one at a time, which is why they are so quick.

Here is our opinion. A large share of the AI features we are asked to add to client websites are not generation problems at all. They are sorting problems: is this enquiry sales or support, is this image safe to show, which of four answers fits this question. Sending those to a frontier chat model is slow, costly and harder to test.

A 39 ms classifier you can run yourself changes the sums. You get a predictable answer shape, a cheap call on every request, and something you can evaluate against your own examples. Cloudflare is also launching a reinforcement learning service so teams can fine-tune on their own data, though the self-serve version is still to come and pricing was not disclosed.

Our advice: before you wire a chatbot into a form, ask whether a decision model could do the job instead.