Cloudflare Clef
Cloudflare Clef is a decision model for developers classifying records, routing requests and choosing the next step in an AI workflow.
Give it the situation and the available answers
Clef evaluates a state, such as a support message or a structured invoice, against questions you define. It returns probabilities for the allowed answers. It does not write a free-form explanation that your application then has to parse into a decision.
For example, a checkout outage can be evaluated for urgency, the team that should handle it, and severity in one request. Your application still decides what to do with the result: create a ticket, send it to technical support, or request human review. The model provides the classification; it does not itself perform those actions.
Three question types, one response
- True or false: the API calls this type
noul. A question such as whether a message expresses urgency returns the probability of true. - Choose a category:
choicetakes named options with descriptions, such as billing, technical support and sales. The response includes the chosen option, confidence and per-option probabilities. - Rate an ordered scale:
scoretakes ordered descriptions, such as no impact, minor, major and critical, and returns a probability-weighted score with its legend.
The Workers AI schema allows one to 64 questions, with answers keyed by the question IDs you supplied. That makes it possible to classify several attributes of the same record in one call. Useful tasks include customer-service routing, invoice processing, tool selection and content categorization. Probability output still needs evaluation on examples from your own workflow.
Clef and Clef-flash
The two models share the decision interface but differ in size and hosted limits. Clef is a 27B model post-trained from Qwen3.8-27B. Clef-flash is a smaller 9B model based on Qwen3.5-9B. The model cards describe a joint schema head that scores the available options, rather than an ordinary text-generation loop.
- Clef on Workers AI:
@cf/cloudflare/clef, with a documented context window of 65,536 tokens and a rate of US$0.24 per million input tokens. - Clef-flash on Workers AI:
@cf/cloudflare/clef-flash, with a documented context window of 24,576 tokens and a rate of US$0.038 per million input tokens. - The model pages say output tokens are not charged. Media is converted into input tokens and charged at the model's input rate.
These rates were checked on October 13, 2026. They are token rates for hosted inference, not the complete cost of a workflow that also fetches pages, runs browser sessions or stores results. Clef-flash is a candidate when latency and cost matter, but the larger model and the smaller one should be compared on the actual categories you need.
Hosted inference and downloadable weights
You can call the hosted models through a Workers AI binding or the REST endpoint using your Cloudflare account and authorization token. The official Clef model card also provides downloadable weights under Apache-2.0 and a local inference example. Its tested setup uses PyTorch and Transformers on an H200; that does not establish a minimum hardware requirement or performance on a consumer GPU.
The model card describes image and video inputs. The current Workers AI request schema explicitly lists embedded PNG, JPEG or WebP images, with at most four images and limits on individual and total size; remote image URLs are not accepted. It does not list a video field. Check the interface you plan to use before assuming the local model's media examples apply unchanged to hosted requests.
Hosted media tokens and questions count toward the context window. The documentation says a request fails if these exceed the window, and otherwise long text state is truncated to fit the remaining space. For a long record, confirm that the evidence needed for the decision is still present.
How to read the comparisons with Jev
Cloudflare describes the request format as Jev/SystemOne compatible, using state and typed questions. That provides a practical starting point for comparing providers with the same task definitions, while endpoint, authentication and model selection still depend on the deployment.
The product image is Cloudflare's own Decision Index score-versus-latency chart, not a Seailife benchmark. The official model card includes internal benchmark and workflow results. Clef leads some rows, Clef-flash leads others, and Jev remains stronger on some tests. A higher aggregate score is not a guarantee of better results for every routing or classification task.
Start with a bounded decision
Cloudflare's launch announcement describes using Clef in a domain-classification workflow with browser rendering, and discusses fine-tuning for support and Trust & Safety use cases. The announcement presents a hands-on customer fine-tuning service, with a self-service platform described as a future direction.
A useful first deployment is one small classification task with labeled examples, an explicit fallback for uncertain answers and a way to measure mistakes. Seailife has not run inference or independently verified the published performance results.