Meet Cloudflare Clef-omni: A 30B-A3B Multimodal Decision Model for Audio, Video, Image and Text

Cloudflare has added Clef-omni to its Clef family, and the hook is simple: one decision model can score text, images, audio and video without the usual detour through transcription and captioning. The release landed on October 9, 2026, just 8 days after the original Clef and Clef-flash debut.

This is an open-weight model with 30B total parameters and 3B active at a time, built as a mixture-of-experts system. It runs on Workers AI under @cf/cloudflare/clef-omni, and Cloudflare says self-hosting was tested on a single H200 with about 64 GB of GPU memory in bfloat16. The context window on Workers AI is 64,000 tokens.

Under the hood, Clef-omni comes from Qwen3-Omni-30B-A3B-Instruct. Cloudflare kept the comprehension backbone, removed the text-to-speech output path, froze the Qwen3 backbone and trained LoRA adapters on top. The model takes a single prefill pass over the full payload, with synced audio and video frames feeding one sequence before a joint schema head scores the options.

The practical pitch is clear enough. Images can go up to 4 per request, audio up to 4 clips, video up to 2 clips, and questions can run from 1 to 64 per request. Cloudflare says text decisions land in about 130 ms median, images in about 150 ms, and a 21-second video with sound scores in about 1.5 seconds.

The benchmark picture is more uneven. Clef-omni does best on intent work, hitting 94.8 macro-F1 on BANKING77 and 97.7 on CLINC150+OOS, and it also tops Amazon ESCI at 57.8. But it lags badly on some harder tasks: 69.3 on home appliances case-exact accuracy versus 97.73 for Clef-flash, 63.3 on When2Call versus Jev’s 80.97, and 47.4 on GPQA Diamond versus Jev’s 78.3.

Cloudflare also used the same update to speed up hosted Clef by moving it to SGLang. The company says median latency improved from 262 ms to 152 ms at about 800 tokens, from 616 ms to 305 ms at about 3,400 tokens, and from 2,721 ms to 1,635 ms at about 16,000 tokens. Clef-flash pricing was cut below TypeSafe’s Jev as part of the broader refresh.