Decision models, unlike large language models, return structured decisions rather than generated text. Until now they have been largely text-only, with Clef supporting images and video frame arrays. Clef-omni extends this to audio in wav or mp3 form and video in mp4 or webm, so developers can call one model across modalities instead of chaining together transcription or channel-splitting pipelines.

Under the hood, Clef-omni is built on a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation, keeping the comprehension backbone while dropping text-to-speech output components. Cloudflare froze the backbone and trained low-rank adapters using the same post-training approach as the original Clef. Because the model skips output token generation, incoming media map directly into a unified sequence, with audio and video synchronised for joint processing.

Speed figures from Cloudflare put median text-only decisions at about 130 ms and image inputs at about 150 ms, with audio clips scored in a few hundred milliseconds and a full 21-second video clip with sound in about 1.5 seconds, all in a single API call. Published benchmarks show Clef-omni performing close to its siblings across tool-calling and classification tests, though it trails Clef on several, such as the Home appliances case-exact score of 69.3 against Clef's 82.95.

On pricing, Clef-flash drops from $0.09 to $0.038 per million input tokens, while Clef remains at $0.24 per million. Cloudflare's post did not state Clef-omni's price, pointing developers to the documentation for current pricing and for details on how image and audio inputs convert to billable tokens. The models are available through Cloudflare's developer docs, with Clef-omni's open weights released on Hugging Face.