Meet Cloudflare Clef-omni: A 30B-A3B Multimodal Decision Model for Audio, Video, Image and Text

Cloudflare released Clef-omni on October 9, 2026, eight days after Clef and Clef-flash. It is a 30 billion parameter mixture-of-experts model with 3 billion active parameters, post-trained from Qwen3-Omni-30B-A3B-Instruct, and it returns typed answers with probabilities instead of free-form text. One call can score across text, images, audio, and video.

That matters for agent pipelines that today stitch transcription and captioning before a text model decides anything. Removing those hops cuts latency and failure points. The same update cut Clef-flash pricing below TypeSafe's Jev and made hosted Clef roughly twice as fast at the median after a move to SGLang.

We like the product shape. Decision models — structured questions that return probabilities over allowed options — are a better control surface for agents than chat. Cloudflare supports three question types (true/false, choice, and score) in a format that follows TypeSafe's Jev and SystemOne API. That is infrastructure thinking, not demo theater.

Clef-omni keeps Qwen3's comprehension backbone, drops the text-to-speech path, freezes the backbone, and trains LoRA adapters with label-smoothed cross-entropy and Brier score calibration. Video frames and audio sync into one sequence. On Workers AI it takes up to four images, four audio clips of five minutes each, two sixty-second videos at two frames per second, and one to sixty-four questions per request, with a 64,000-token context. Cloudflare reports text decisions around 130 ms median, images around 150 ms, and a twenty-one-second video with sound in about 1.5 seconds. Hosted price is $0.15 per million tokens under Apache-2.0, with self-hosting tested on a single H200 at about 64 GB in bfloat16.

I'm watching the accuracy trade. Clef-omni leads the Clef family and Jev on intent tasks: 94.8 macro-F1 on BANKING77, 97.7 on CLINC150+OOS, 57.8 on Amazon ESCI. It trails on harder work — 69.3 case-exact on home appliances versus 97.73 for Clef-flash, 63.3 on When2Call versus 80.97 for Jev, and below Clef on all five reported TypeSafe workflow metrics. Worst published marks include 2.6 accuracy on POP909-CL and 47.4 on GPQA Diamond against Jev's 78.3. Multimodal convenience is not free. If teams route hard reasoning through omni because the API is prettier, they will ship worse decisions.

The dual-use angle is quieter here than in cyber models, but it is real. A calibrated probability over "approve / reject / escalate" across live audio and video is powerful for moderation, fraud, and access control. Useful when the schema and thresholds are explicit and auditable. Harmful when a low-confidence multimodal score silently gates people without a human review path. Decision models make that choice visible — which is why we prefer them to opaque chat agents — but visibility only helps if operators actually read the confidence numbers.

Our read: use Clef-omni for multimodal intent and routing, keep text-focused Clef or Jev for hard reasoning and workflows, and treat the probability outputs as something to log and threshold, not something to trust blindly. Useful AI at the edge of the network looks like this when the evals stay honest.

MarkTechPost notes that benchmark scores come from Cloudflare's Decision Index announcement and were not independently verified. Hosted Clef's median latency fell from 262 ms to 152 ms around 800 tokens and from 616 ms to 305 ms around 3,400 tokens after the SGLang move, with no new Clef weights. Clef-flash now lists at $0.038 per million input tokens versus Jev's $0.042.