Cloudflare: Clef

Cloudflare Clef is the larger dense model in the Clef decision family, post-trained from Qwen3.8-27B with its vision encoder. Cloudflare combines that backbone with a joint schema head that routes evidence to the questions and scores their allowed options together. Clef reads a situation and returns bounded decisions with probabilities in a single forward pass. Its output is a classification result, rather than generated prose, an open-ended report or a proposed command.

Prepare the state and question schema as distinct inputs. The state can be a string or JSON representing a record, conversation, document content or application situation. Questions are keyed by stable names, and the same names identify their returned answers. Define what each question should establish and describe the relevant options. A support record can ask which team should handle the request, whether the impact is urgent and which severity criterion applies, without combining those different decisions into one vague category.

A noul question returns the probability of true for a binary condition. A choice question returns a selected named option, per-option probabilities and confidence. A score question uses an ordered list of descriptions, indexed from zero, and returns an expected score, legend and probabilities. A fractional score is meaningful because it summarizes uncertainty across ordered levels. Inspect the distribution as well as the selected answer; two close options can reveal a distinction the input does not settle.

For policy checks, include the actual rule text and the facts of the record. An example is a document-processing workflow with questions for whether the record is complete, which review category applies and how strongly the evidence supports immediate attention. Ask separate questions for separate conditions. A model assessment of compliance should remain connected to the supplied rule and evidence. If an important condition is missing, the schema and application should have a deliberate way to handle that uncertainty.

For visual evidence, provide a readable image and a question about a visible condition or a defined category. A form-quality check might evaluate legibility, whether a required area is present and which of the supplied quality levels best describes it. The maker’s open-weight interface also accepts video frame arrays. Keep the frames in order and identify the event to assess. A bounded decision schema can classify that material, while a request for a narrative description or arbitrary newly generated fields belongs to a different kind of output.

The manufacturer hosted reference lists a 65,536-token context, one to 64 questions and up to four embedded images. Its PNG, JPEG or WebP images are limited to 4 MiB and 16 megapixels each, with 8 MiB total decoded; the whole request body is limited to 13 MiB, and remote image links are not accepted. The input budget includes the questions and media. Long text may be truncated to fit, so preserve decisive passages and source labels. These limits describe the producer’s hosted image interface; the open-weight encoder has its own local-input representations.

Cloudflare’s release describes training for schema-bound scoring and probability calibration. The model extracts relevant evidence for individual options and then considers information across fields. This can support decisions over several related questions in one record, but it does not make a high probability a guarantee that the underlying fact is true. Evaluate difficult, incomplete and out-of-category records as well as straightforward cases. Useful decision criteria cover the situations the workflow actually encounters and make important distinctions explicit.

For an agent workflow, the output can determine which defined next step should be considered. Supply the current observation and candidate actions with clear descriptions, then retain the selected option and its probabilities. The surrounding application applies its action policy and records what it did. Keep that execution record separate from the decision result: selecting a review route does not establish that a reviewer acted, and selecting a diagnostic step does not establish that the step succeeded.

The final artifact is the answers object together with the state and schema that produced it. Use those records to inspect unexpected classifications, identify ambiguous criteria and compare results with reviewed task examples. If a downstream user needs a written explanation, the decision and original evidence can be inputs to a separate response-writing step. Clef supplies the bounded probabilistic decision needed to choose that next stage with a clear account of what was evaluated.