Cloudflare launches Clef, a decision model rival to Jev

Cloudflare has shipped its first in-house machine-learning models, a pair of decision models called Clef and Clef-flash, hosted on the company's Workers AI platform. It has also opened a reinforcement-learning service that lets customers fine-tune Clef on their own data.

A decision model is not a chatbot, and it never writes prose. You hand it an input, whether that is a customer support message, a website domain or a photograph, and it returns a bounded, typed answer with a probability attached, such as an urgency flag at 0.95 confidence or a billing category. The point is to give an autonomous agent a cheap, fast way to choose its next move.

What a decision model actually does

Cloudflare's own example is a support ticket: pass one in and ask which team should handle it and whether it is urgent. Clef answers with typed values and confidence scores rather than a paragraph of text, which is something an agent can act on directly rather than parse.

The company has been testing Clef internally on its Threat Intelligence team, where the model classifies website domains, working alongside Browser Run to sort a domain into categories with a confidence figure attached.

Picking a fight with Jev

Clef arrives in a market that got loud, fast. Typesafe AI's Jev System One kicked off a wave of interest in decision models, and Cloudflare is not being subtle about the comparison. It says Clef currently leads on the Jev Decision Index, and that the model is fully API compatible with Jev, so an existing integration can swap models without a rewrite.

Two differences it is keen to press. Clef carries a vision encoder, so it can classify images as well as text, which Cloudflare says Jev cannot do. And on Typesafe's own evaluation suite, Cloudflare claims Clef beat Jev in three of the four areas measured.

Latency is the other pitch. Across 43 evaluation benchmarks, Cloudflare says its models came out ahead on speed against every rival bar Laya, which it concedes is faster but gives up quality to get there. Because Clef runs on Workers AI, inference happens on GPUs at Cloudflare's edge rather than in a distant region.

How Clef was built

Under the hood, Clef is a post-trained Qwen model. The full-size Clef freezes Qwen3.8-27B while Clef-flash uses Qwen3.5-9B. Both backbones are frozen, with a routing head trained jointly alongside rank-256 low-rank adapters. Training paired label-smoothed cross-entropy for valid schema outputs with a Brier loss to sharpen probability calibration. At inference, Clef runs a prefill-only pass and then scores the valid choices in parallel.

Fine-tuning, and the new RL service

The pitch to businesses is specificity. A general classifier is fine until you need one that understands your own taxonomy, and Cloudflare's argument is that labelled decisions pile up quickly inside a large company. Its fine-tuning offer starts hands-on, with Cloudflare's forward-deployed engineers doing the work, and is meant to become a self-serve platform where customers capture data, fine-tune and redeploy the model.

The reinforcement-learning product is assembled from pieces Cloudflare already sells. AI Gateway captures AI traffic and harvests request and response data, Containers run the RL sandboxes, and Workers AI's bring-your-own-model work, codenamed Cog, supplies the training stack. The model weights are published on Hugging Face for anyone who wants to poke at them.

Our opinion

Decision models are the least glamorous corner of AI and probably the one that will end up doing the most work. Nobody demos a classifier on a stage. But every agent that has to choose between five possible next actions needs something cheap and fast to make that call, and using a full language model for it is like hiring a novelist to write your shopping list. Cloudflare is right that there is a real market here.

The interesting part is not the benchmark spat with Jev, which is a fight over a leaderboard Typesafe itself set up. It is the fine-tuning pipeline. Cloudflare already sees a customer's AI traffic through AI Gateway and can run the training job in its own containers, which means it can offer a decision model trained on a customer's actual decisions without that data leaving the building it was already sitting in. That is a stickier product than a clever model, and it is the piece rivals without the same plumbing will struggle to copy.