Cloudflare has released Clef and Clef-Flash, open-weight decision models. Both models take a state, such as text, JSON, or images, plus a question, and return probabilities for a fixed set of options. Cloudflare hosts the models on Workers AI and has released the weights on Hugging Face under the Apache 2.0 license.
Clef returns probabilities for fixed options
Instead of continuing a free-form answer like an ordinary language model, Clef returns a structured judgment an agent needs to pick its next action. A request carries the current state and one or more questions, and each question is expressed as one of these three types.
noul: Judges true or false, like yes or no.choice: Judges which of the given options to pick.score: Produces a score for an ordered value.
The response includes a probability for each answer. You can put several questions in one request, which suits agent steps where options can be defined ahead of time, like classification, routing, and policy decisions. Cloudflare's Clef announcement presents it as a model for making fast, repeatable decisions in agent workflows.
Clef and Clef-Flash: a choice between accuracy and latency
The two models have different sizes and latency profiles. Cloudflare's model cards describe Clef as a 27B model based on Qwen3.8-27B and Clef-Flash as a 9B model based on Qwen3.5-9B.
| Model | Base model and size | Recommended use per Cloudflare | Median latency | p95 latency |
|---|---|---|---|---|
| Clef | Qwen3.8-27B · 27B | Accuracy-first judgments | 209.3ms | 238.6ms |
| Clef-Flash | Qwen3.5-9B · 9B | Latency-sensitive hot paths | 38.8ms | 122.4ms |
| Jev | System One decision model | Comparison baseline | 524.1ms | 536.0ms |
These figures are latencies from 43 evaluation runs reported by Cloudflare. The announcement's "13x faster than Jev" comparison corresponds to Clef-Flash's median of 38.8ms versus Jev's 524.1ms. At p95, they are 122.4ms and 536.0ms respectively. The median and p95 are different measures, so it pays to keep them apart. The comparison describes that Cloudflare evaluation's results and is not presented as a universal speed figure for every input or runtime environment. You can see the details in the Workers AI changelog and the Clef model card.
Workers AI hosting and open weights
Cloudflare has made both models available on Workers AI, and it provides the weights under Apache 2.0 in its Hugging Face repositories. Hosting works by sending the state and questions as a Workers AI request. In Cloudflare's example, you call it from a Workers environment in the form env.AI.run("@cf/cloudflare/clef", { model: "clef", state, questions }).
The model card also describes a way to run the downloaded weights yourself. The Clef-Flash card has a Python example that downloads the model files and handles requests with load_release_model and the systemone helper. The verification environment the card lists is PyTorch 2.11, Transformers 5.10.2, and a single NVIDIA H200. That is the environment the example was verified in and doesn't indicate the minimum memory spec for ordinary local runs. The model card gives no minimum VRAM or RAM figure.
The model card also describes an interface that includes arrays of images and video frames in the input. In the local Python example, the images or frames are passed to local preprocessing code. The M5 Max webcam demo shown in the announcement is also a local inference demonstration. If you choose Workers AI hosting, by contrast, your input requests are sent to Cloudflare's hosted service. Data handling differs between the two deployment paths.
Request and response formats compatible with Jev
The systemone helper in the Clef model card is compatible with the request body format and response structure of Jev/System One's POST /v1/systemone. The response includes model, answers, usage, and more, and expresses noul, choice, and score results. Compatible here means within the scope of that API's request and response schema. It doesn't mean it is the same model or service as the whole Jev product.
To explore decision models and the Jev ecosystem further, read the introduction to CLM-8B and the System One decision model Jev, Unsloth Laya's Jev-compatible local run, and OpenAI Decisions API.
RL fine-tuning is currently offered as a custom service
Along with the Clef announcement, Cloudflare also revealed plans for a reinforcement learning (RL) fine-tuning platform for decision models. What it offers now is a service in which Cloudflare's FDE (forward-deployed engineering) team works with you to tune the model to your workload, and it is accepting design partner applications. A self-service platform where customers collect their own data, fine-tune, and redeploy is a future plan. So you should distinguish between what customers can use self-service today and the platform to be built later.