Cloudflare releases Clef: open-source decision models and a fine-tuning platform

Cloudflare has released two of its own decision models, Clef and Clef-flash. Both are open-sourced under the Apache 2.0 license on Hugging Face, hosted on Workers AI, and fully compatible with the API of Typesafe AI's Jev System One — the model that has generated a lot of buzz in recent weeks.

What a decision model is

Unlike large language models, which are non-deterministic and can reason and generate free-form text, a decision model returns a bounded, strictly typed output quickly, cheaply and consistently. You feed it inputs and a set of questions, and it returns answers with probabilities: given a support message, for instance, it can decide whether it is urgent, which team should handle it, and how severe the customer impact is. Your code can then use that to route, escalate, or defer to a human.

Cloudflare already uses Clef in its Threat Intelligence team to classify domains. The company says Clef categorised a website (with probabilities — fashion, e-commerce, phishing and so on) in 2.2 seconds, whereas its fastest general-purpose LLM, gpt-oss-120b, took 4.7 seconds and returned only two categories.

How Clef differs

Clef has a vision encoder, so it can classify images — Jev currently handles text only. Its context window is 64k, against Jev's 32k. On the Jev Decision Index tables and Typesafe's own eval suite, Cloudflare's models score competitively, and on median latency they outpace most rivals.

Clef is built on Qwen, post-trained for decision-making tasks (Qwen3.8-27B for Clef, Qwen3.5-9B for Clef-flash). The decision step is non-autoregressive: no intermediate text is generated, and schema choices are scored in parallel, which is where the speed gain comes from. To calibrate probabilities, Cloudflare developed its own reinforcement-learning method, RLCD (Reinforcement Learning for Calibrated Decisions).

Fine-tuning platform

Alongside the models, Cloudflare is launching a service to fine-tune Clef for specific workloads — first hands-on with its forward-deployed engineering (FDE) team, and later as a self-serve platform that captures data, trains, and redeploys the model entirely on Cloudflare's infrastructure. It builds on existing pieces: AI Gateway (turning traffic into a dataset), Workers AI, Containers (as an RL sandbox) and a new Trainer component, plus the Bring Your Own Model capability that grew out of the Replicate acquisition.

The name nods to the musical clef that fixes the pitch of notes on a staff — while the CF hints at Cloudflare.

Source: Cloudflare Blog