Model overview
What this model is for
Cloudflare Clef-Omni is an Apache-2.0 typed-decision model built on the Qwen3-Omni-30B-A3B-Instruct backbone. It takes text or JSON state plus images, audio, and video, and scores caller-defined choice, yes/no, and ordered-score questions with per-option probabilities instead of free-form prose. The architecture is 30B-A3B (about 3B active per pass), while Hugging Face reports roughly 35B parameters across repository files because unused speech-output weights are present but not loaded by the release loader. The hosted model lists a 64K context; media tokens count toward it. The model card reports about 64 GB GPU memory for BF16 self-hosting, so the 3B active count is not a hardware-fit estimate. It fits multimodal triage and routing that require a fixed output schema, not transcription or general conversation. Integrate via the model-specific SystemOne interface, enforce permissions and human review outside the model, and test task-specific calibration before acting.