Clef-omni

Clef-omni is the multimodal member of Cloudflare's Clef family of decision models, released on October 9, 2026. It is post-trained from Alibaba's Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts with 30B total and 3B active parameters, keeping its vision and audio encoders and adding the joint schema head the Clef models share. You send a state (text, any JSON value, images, audio or video) and named questions, and it returns typed answers: the probability that a statement is true, a choice among options you define, or a position on your own rubric, each with a full probability distribution. It generates no text, so answers cannot fall outside the options given. A video's soundtrack is aligned with its frames, so one request can weigh what is seen and heard together where a text decision model would first need a transcript and extracted frames. Cloudflare reports about 20 ms for a text request, under 100 ms with an image or audio input and about 300 ms for a 21-second video with sound. It speaks the same request and response format as TypeSafe's System One API and the other Clef models, so code written for Jev, Clef or Clef-flash can switch by changing the model name. Use it for decisions about calls, recordings, photos and clips, such as inspection checks, content review or triage of voice and screen recordings. Media are converted to input tokens and count toward the 64,000-token window together with the questions: a request whose media exceed it fails, and a long text state is truncated to fit. Each request takes up to four audio clips and two videos. Weights are published by Cloudflare under Apache-2.0 at https://huggingface.co/Cloudflare/clef-omni.