Cloudflare says its new Clef-omni model can score audio alongside text, images and video in a single request, returning probabilities for application-defined answers instead of an open-ended written response.
Photo by Stephan Ridgway via Wikimedia Commons, licensed under CC BY 2.0.
The company’s example combines a photo of an appliance, a recording of it running and a video of its fan. An application can ask whether the model label is visible, whether the sound seems normal and whether the fan is turning. Cloudflare says a single model can handle those inputs together instead of sending them through separate transcription and image-analysis steps.
The change extends Cloudflare’s Clef models, which already accepted text, images and video frames. Clef-omni adds audio and synchronized audio-video processing. It is a decision model, not a chat model: an application defines yes-or-no questions, choices or ordered scores, and the model returns the allowed answers with probabilities. Cloudflare says it uses Qwen3-Omni’s comprehension backbone but drops speech generation, scoring the input against valid answer options in one pass. The model card says video files are sampled at two frames per second and their soundtracks are processed with the frames when present.
Cloudflare says the hosted model is available through Workers AI and has released its weights under Apache 2.0. The Hugging Face model card lists 35 billion parameters and says it needed about 64 GB of GPU memory for its bfloat16 test on a single H200. That makes the weights reusable, but local deployment still calls for substantial hardware.
Cloudflare’s own benchmark table gives mixed results against the earlier Clef. Clef-omni scores 69.3 on the home-appliance exact-match test, below Clef’s 82.95, while its BANKING77 macro-F1 is 94.8 versus Clef’s 94.20. The announcement presents these as company evaluations, not an independent comparison.