Clef-omni

Clef-omni is a decision model from Cloudflare, released on 9 October 2026 alongside a faster Clef and a cheaper Clef-flash. Like the rest of the family it does not write free text. It reads a state described in text or JSON, optionally with images, audio and video, together with a set of questions, and returns one score for each allowed option, so software can act on a bounded, structured answer. Cloudflare says it is post-trained from Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts model of 30B parameters with about 3B active, with the speech-output parts left unused. The weights are on Hugging Face under the Apache 2.0 licence, where the card also notes a single GPU with about 64 GB of memory for the backbone in bfloat16. The model runs on Cloudflare's Workers AI at a launch price of $0.15 per million input tokens. The post gives median latencies of about 130 ms for text and about 150 ms for an image. The accuracy results on the model card come from Cloudflare's own internal evaluation, and the card shows the larger Clef and a rival decision model ahead on several of them.

Not yet independently verified: Every detail here comes from Cloudflare's own blog post of 9 October 2026 and the model card on Hugging Face; no independent report was found. The accuracy figures on the card are from Cloudflare's own internal evaluation, so no score is stored. The blog post names no licence; the Apache 2.0 licence is taken from the model card. We will update this when it can be confirmed, and remove this note.