Cloudflare: Clef Omni

Cloudflare Clef Omni is the multimodal mixture-of-experts model in the Clef decision family. It is post-trained from the Qwen3-Omni-30B-A3B-Instruct comprehension backbone, with 30B total and 3B active parameters. Cloudflare adds a joint schema head that routes evidence to the requested questions and scores their allowed options together. The result is a typed decision with probabilities, rather than a generated explanation, transcript or speech output.

Prepare two main parts: a state describing the situation and a mapping of named questions. The state can be text or structured JSON, and the model’s native inputs also include images, audio and video. Each question says what to decide and which answers are allowed. Keep identifiers stable so an application can associate each returned answer with the original question. Clear option descriptions make a category meaningful; a short ambiguous label can otherwise hide the distinction the workflow actually needs.

The record format has three question types. A noul question asks about a true/false condition and returns the probability of true. A choice question defines named options and returns a selected option, per-option probabilities and a separate confidence value. A score question defines ordered criteria indexed from zero and returns a probability-weighted score, a legend and probabilities. The score may be fractional because it reflects a distribution over levels. A decision is therefore richer than a single label, while still being bounded by the supplied schema.

For visual review, provide a clear image and the condition to assess. An example is an installation record with questions asking whether the identification label is visible, which of the defined condition categories applies, and whether the evidence is sufficient for review. Keep observations separate: a label being visible is a different question from an installation appearing complete. The output can help an application route the record or flag a review condition; it does not itself perform an inspection, change a record or generate a written report.

For audio and video, focus the questions on evidence the supplied clip can support. A practical equipment review might include a photo, a recording and a short video, asking whether a specified sound is present or whether a component is visibly moving. The maker describes processing a video’s soundtrack alongside its frames when every video in the record has sound. Preserve the relationship between the media and the written state, and avoid asking the model to classify an event that lies outside the attached interval.

The manufacturer’s hosted profile lists a 64,000-token context and one to 64 questions. It accepts up to four images, four audio clips and two videos, with audio clips limited to 300 seconds each and videos to 60 seconds each. Those are limits of the documented producer interface. Media and questions occupy context; excessive media can fail the request, while a long text state may be truncated to fit. Put decisive evidence in a focused record and prepare the questions before spending the window on unrelated history.

The maker’s hosted interface takes embedded media rather than remote links. Its PNG, JPEG or WebP images are limited to 4 MiB and 16 megapixels each, with an 8 MiB total decoded image budget. Each audio clip is limited to 8 MiB and 300 seconds; each video to 16 MiB and 60 seconds. Audio and video together may total at most 16 MiB decoded, and videos are sampled at two frames per second. These limits belong to that producer interface; the open-weight encoder accepts additional local representations. The comprehension backbone’s audio capability does not mean this decision model produces audio: its speech-generation components are not used for decision scoring.

For support or workflow routing, define categories and ordered criteria from the actual operating rules. A ticket record might ask which team should handle it, whether the impact is urgent and which severity level best matches the evidence. The observable output is a set of typed answers under those question names. Review the whole probability distribution when categories are close, and define the application’s treatment of uncertain cases explicitly. A high probability is a model assessment of the supplied evidence, not proof that the condition was independently verified.

Retain the state, the question schema and returned answers together when inspecting a result. Check whether the options cover the actual situation, whether their descriptions overlap, and whether important evidence was omitted or truncated. If a downstream step needs a written explanation, it can use the decision and source material as input to a separate writing step. Clef Omni’s useful role is scoring defined alternatives from multimodal evidence so the next action can be chosen with a clear, inspectable input and output.