clef-omni

Overview. Clef-Omni is an open-weight multimodal decision model from Cloudflare, released under the Apache-2.0 licence. It is post-trained from Qwen3-Omni-30B-A3B-Instruct and belongs to Cloudflare's Clef family. It scores a set of predefined options and returns a probability for each one. It does not generate free-form text. Reach for it when you want a structured decision from mixed-media input.

Key strengths. Clef-Omni accepts text, JSON, images, audio and video as input. Its mixture-of-experts backbone has a nominal 30B parameters, with about 3B active per token. The backbone configuration lists 128 experts, with 8 routed per token. Cloudflare's model card says option logits and probabilities come out of a single forward pass through a custom joint schema head, which gives you structured decisions without a generation loop. Cloudflare's model card does publish benchmark scores for Clef-Omni, including classification accuracy on BANKING77 and CLINC150+OOS and TypeSafe Evals workflow results, alongside the architecture details.

When to use it. Use Clef-Omni for classification-style decisions where the input is an image, an audio clip, a video or a JSON payload and the answer is one of a known set of options. Examples are routing, triage and labelling. It also suits cases where you want probabilities you can threshold, rather than text you have to parse. The weights are public under Apache-2.0, so you can run it yourself, and Cloudflare also hosts it on Workers AI.

Limitations. Clef-Omni is not a general chat or text-generation model. It returns option probabilities, not free-form completions, and it produces no image or audio output. The hosted Workers AI endpoint documents a 64,000-token context window, slightly below the 65,536 positions in the released backbone configuration. No training-data cutoff is stated on the model card.