Cloudflare has released Clef-omni, a new open-weight member of its Clef decision model family. Clef-omni is a Cloudflare decision model that scores text, images, audio and video in 1 call. It arrived on October 9, 2026, 8 days after the original Clef and Clef-flash launch. The same update cut Clef-flash pricing below TypeSafe’s Jev and made hosted Clef roughly 2x faster. For developers, this removes the transcription and captioning steps that multimodal agent pipelines usually need.
TL;DR
- Size: 30B total, 3B active parameters (MoE); 64,000-token context on Workers AI
- Runs on: Workers AI as
@cf/cloudflare/clef-omni; self-hosting tested on a single H200, about 64 GB GPU memory in bfloat16 - Performance: Close to Clef on routing and intent tasks, behind it on most reasoning and workflow evals
- Best: 97.7 macro-F1 on CLINC150+OOS, top of the 4 Clef and Jev models
- Worst: 2.6 accuracy on POP909-CL, and 47.4 on GPQA Diamond versus Jev’s 78.3
- Bottom line:
- Best: one call decides across 4 modalities.
- Worst: weaker accuracy than text-focused Clef on hard reasoning.
What is a decision model?
A decision model returns typed answers with probabilities instead of generated text. You send a state and a schema of questions. The model returns a probability for every allowed option. Clef supports 3 question types: noul (true or false), choice and score. The format follows the Jev and SystemOne API from TypeSafe AI.
How does Clef-omni work?
Clef-omni is post-trained from Qwen3-Omni-30B-A3B-Instruct. Cloudflare keeps the comprehension backbone and drops the text-to-speech output path. The model runs a single prefill pass over the full payload. Video frames and audio are synced into 1 sequence for joint processing.
A small joint schema head then scores the options. Each valid option first gathers evidence from the input. Field vectors then cross-attend across the full context to compute confidence scores. Cloudflare froze the Qwen3 backbone and trained LoRA adapters. Training combined label-smoothed cross-entropy with Brier score calibration.
Supported inputs on Workers AI, per the developer docs:
- Images: up to 4, PNG, JPEG or WebP, 64 to 1,024 tokens each
- Audio: up to 4 clips, WAV or MP3, 300 seconds each, about 780 tokens per minute
- Video: up to 2 clips, MP4 or WebM, 60 seconds each, sampled at 2 frames per second
- Questions: 1 to 64 per request
Cloudflare reports text decisions in about 130 ms at the median and images in about 150 ms. A 21-second video with sound scores in about 1.5 seconds.
How does Clef-omni perform on benchmarks?
Cloudflare compared Clef-omni against Clef, Clef-flash and Jev on its Decision Index run. Clef-omni leads on intent classification: 94.8 macro-F1 on BANKING77 and 97.7 on CLINC150+OOS. It also tops Amazon ESCI at 57.8.
It trails elsewhere. Home appliances case-exact accuracy is 69.3, versus 97.73 for Clef-flash. When2Call accuracy is 63.3, versus 80.97 for Jev. On TypeSafe’s workflow evals, Clef-omni scores below Clef in all 5 reported metrics.
