Cloudflare open-sources Clef decision models and launches an RL fine-tuning service

Cloudflare introduced two decision models, Clef and Clef-flash, on October 1, 2026. According to the company, they are the first models trained in-house by its Workers AI team. Both run hosted on Workers AI, Cloudflare's AI platform, and the open weights are available on Hugging Face under the Apache 2.0 license, so organizations can run them on their own servers. In the same announcement, Cloudflare launched a fine-tuning service based on reinforcement learning (RL) that adapts Clef to a customer's own workload.

Decision models assign probabilities to options

A decision model is built for the moments inside a workflow where software or an AI agent needs to make a quick call. You give it a state, such as a customer support message, and questions with predefined answers: is the request urgent, which team should handle it (billing, technical or sales), how severe is the impact on the customer. The model returns a probability for each option without writing a long answer. Your code can then route the ticket, escalate it, or hand the decision to a person when confidence is low. Because the options are defined in each request, adding a new category does not require retraining.

Cloudflare notes that this kind of model drew attention in recent weeks with Typesafe AI's Jev. Clef's request and response format is compatible with Jev, so an application that already uses Jev can switch easily. The company says it has been testing Clef in its own threat intelligence team to classify websites: the model scores which categories a domain falls into, including the likelihood that it is a phishing site, in one pass.

Two sizes, image input and a wide context

According to the announcement, the larger Clef is built on Qwen3.8-27B, while the smaller and faster Clef-flash is built on Qwen3.5-9B. Cloudflare kept the base Qwen models frozen and added a small head that scores the options, along with extra trained adapter layers. The model reads the state once and scores all options together; since it does not generate text token by token, it runs noticeably faster than general-purpose language models. According to the Hugging Face pages, both models accept text, JSON, images and video frames as input. The announcement gives a 64K-token context window. Cloudflare recommends the larger model for precision work and Clef-flash for latency-critical decisions.

In the comparisons Cloudflare published, Clef ranks first on the Jev Decision Index and beats Jev in most of Typesafe's own workflow tests, while rival models lead in some others. The results are also available on a live leaderboard set up by Cloudflare.

Availability and the fine-tuning service

On Workers AI, the models can be called as @cf/cloudflare/clef. Cloudflare says it does not read, store or train on these requests, with separate terms for customers who opt into the fine-tuning service. Anyone who wants to try the models on their own hardware can download them from Hugging Face, where the model cards include Python usage examples. Current prices are on Cloudflare's site.

The new RL service arrives in two phases. First, Cloudflare's forward-deployed engineering team works hands-on with a customer to fine-tune Clef for that customer's workload. Later, a self-serve platform is planned where customers can capture their own data, fine-tune the model and redeploy it on Cloudflare; no date has been given for that phase. The system brings together existing Cloudflare pieces such as AI Gateway to record traffic, Containers for the training sandbox and Workers AI to serve the tuned model. An interest form is open for companies that want to take part.

All the published comparisons were run by Cloudflare. The announcement and model pages do not say which languages the models handle or how well, so it is worth testing them on your own examples before using them on non-English text. Running the larger model yourself calls for a powerful data center GPU; according to the model card, it was tested on a single NVIDIA H200. For a fuller picture, it is worth following evaluations from independent researchers.

Sources: Cloudflare: Introducing Clef decision models and RL fine-tuning, Hugging Face: Cloudflare Clef, Hugging Face: Cloudflare Clef-flash.