Cloudflare open-sources Clef-Omni multimodal decision model on October 10 with native audio and video input

Cloudflare open-sourced the Clef-Omni model weights on Hugging Face on October 10 under Apache-2.0, a multimodal model aimed at structured decision tasks. Clef-Omni uses the Qwen3-Omni-30B-A3B-Instruct thinker backbone with vision and audio encoders, totaling around 35B parameters with 3B activated in the mixture of experts. It takes text, JSON, images, WAV/MP3 audio, and MP4/WebM video as input, and produces the probability of every allowed option across a typed schema in a single forward pass rather than free text. On internal Decision Index 0.2.1, Clef-Omni posts best-of-family scores on MMLU accuracy 92.7 percent, HellaSwag 99.1 percent, ARC-Easy 99.8 percent, and several others; workflow evals include invoice exact-action 60.2, customer service exact-action 71.6, and security exact-action 61.7. The release also cuts Clef-Flash input pricing from $0.09 to $0.038 per million tokens, a 58 percent drop, and reduces the managed model's context window from 64k to 24k. Clef-Omni is API compatible with Jev and SystemOne and is shipped with vLLM, Docker Model Runner, and SGLang support coming soon.