Clef-omni reads video: a cheap check before a second video job

Clef-omni, announced by Cloudflare on October 9, 2026, is an open-weight decision model that reads text, images, audio and video in one request and returns a probability for each allowed answer. Cloudflare prices it at $0.15 per million input tokens and says a 21-second video with sound takes about 1.5 seconds. Sume does not list Clef, so use it outside Sume as a cheap check before you pay for another generation.

What does Clef-omni do?

Cloudflare's post says Clef-omni accepts wav and mp3 audio, mp4 and webm video, images and text in a single API call. It runs a prefill pass over the whole payload and scores every valid option for each question at once, so the output is a set of confidence values, not free text. It is built on Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts model, with the text-to-speech parts removed, and the weights are on Hugging Face.

How fast and how cheap are the three tiers?

The post gives these numbers for the launch. Latency figures are medians Cloudflare reports for short payloads.

Where would a video check fit with Sume?

A Sume video job is asynchronous and billed per output second, so a rejected clip costs a full render. A fast yes-or-no check on an input photo or a finished clip, run in your own code before you submit the next job, can stop obvious misses: the wrong product in frame, text where there should be none, a person you did not intend. Sume does not run Clef for you: the Sume repo has no Clef model, so you would call Cloudflare yourself.

• Ask closed questions with a fixed answer list, such as product visible: yes or no.
• Only submit the next Sume video job when the confidence is above a threshold you choose.
• Keep the threshold in your code; a confidence score is not a spend cap.

What should you verify first?

Confirm context limits and pricing on Cloudflare's model page before you build, because the hosted Clef-flash context differs from the self-hosted weights. Then test the gate on a few dozen of your own clips: a decision model can be confidently wrong on your particular catalog, and the cheapest check is the one you have measured.