Cloudflare Launches Clef-omni Decision Model

Cloudflare has released Clef-omni, an open-weight multimodal decision model under an Apache-2.0 license. Launched on October 9, 2026, eight days after Clef and Clef-flash, it features 30B total parameters with 3B active via a mixture-of-experts architecture. Post-trained from Qwen3-Omni-30B-A3B-Instruct using frozen backbones and LoRA adapters, Clef-omni scores text, images, audio, and video in a single call, dropping transcription steps in agent workflows. Hosted on Workers AI as @cf/cloudflare/clef-omni with a 64,000-token context window at $0.15 per million tokens, it can also be self-hosted on a single H200 GPU with 64 GB GPU memory in bfloat16.

The model accepts up to 4 images (64 to 1,024 tokens in PNG, JPEG, or WebP), 4 audio clips (WAV or MP3 up to 300 seconds, generating about 780 tokens per minute), and 2 video clips (MP4 or WebM up to 60 seconds sampled at 2 frames per second). Following TypeSafe AI’s Jev and SystemOne API, it scores noul, choice, and score questions. Median latencies are 130 ms for text, 150 ms for images, and 1.5 seconds for a 21-second video with audio. Cloudflare also lowered Clef-flash pricing to $0.038 per million tokens—below Jev's $0.042—and sped up hosted Clef (priced at $0.24 per million tokens) by 1.7x to 2.0x via SGLang 0.5.22, reducing median latencies across 800, 3,400, and 16,000 tokens from 262 ms to 152 ms, 616 ms to 305 ms, and 2,721 ms to 1,635 ms.

In benchmarks, Clef-omni led intent classification with a 97.7 macro-F1 on CLINC150+OOS (versus Clef's 97.43, Clef-flash's 66.77, and Jev's 89.27) and 94.8 on BANKING77 (versus Clef's 94.20, Clef-flash's 90.93, and Jev's 79.74), plus 57.8 on Amazon ESCI. For primary action invoice processing, it scored 82.0 (Clef: 86.2, Clef-flash: 73.3, Jev: 83.1). It trailed on BRIGHT nDCG@10 with 42.0 (Clef: 45.91, Clef-flash: 39.26, Jev: 47.52) and When2Call with 63.3 (Clef: 72.37, Clef-flash: 65.58, Jev: 80.97). On specialized reasoning, it recorded 69.3 on home appliances case-exact accuracy (versus Clef-flash's 97.73), 47.4 on GPQA Diamond (versus Jev's 78.3), and 2.6 accuracy on POP909-CL.