Cloudflare adds audio and video with Clef-omni
Cloudflare has released Clef-omni, an open-weight decision model that accepts audio and video alongside text and images, so developers can make multimodal decisions in one API call instead of chaining transcription pipelines.
Cloudflare announced Clef-omni a week after it open-sourced its Clef decision models. The company also said it has made Clef faster and cut the price of Clef-flash so that it now costs less than Jev, TypeSafe's decision model. The post gives no new price or speed figures.
Clef-omni accepts audio (wav or mp3) and video (mp4 or webm) on top of text and images. The original Clef handled images and arrays of video frames. Clef-omni is built on Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts model. Cloudflare kept that model's comprehension backbone and dropped its text-to-speech output parts. Clef models don't generate text. Instead, they score a fixed set of valid answer options in a single prefill pass, which Cloudflare says removes the need to transcribe or caption media first.
Training followed the same recipe as Clef. The Qwen3 backbone was frozen and only LoRA adapters were trained, using label-smoothed cross-entropy combined with Brier score calibration. The weights are available on Hugging Face, and the model also runs on Workers AI.
Cloudflare reports these median latencies:
- Text-only decisions: about 130 ms
- Image inputs: about 150 ms
- Audio clips: a few hundred milliseconds
- A 21-second video clip with sound: about 1.5 seconds
In Cloudflare's own benchmarks, Clef-omni is roughly level with Clef on BFCL (98.2), BANKING77 (94.8) and CLINC150+OOS (97.7). It scores well below Clef on its home-appliances test (69.3 vs. 82.95) and on When2Call (63.3 vs. 72.37). Jev leads all three Clef models on When2Call and BRIGHT. None of these results have been independently verified.