Cloudflare's Clef-omni Scores Text, Audio, Video in One $0.15 Call

Cloudflare released Clef-omni on October 9, 2026, a 30B-parameter mixture-of-experts decision model that scores text, images, audio and video in a single call, priced at $0.15 per million input tokens. The launch came just 8 days after the original Clef and Clef-flash debut, and the same update cut Clef-flash pricing below TypeSafe AI's Jev and made hosted Clef roughly 2x faster. For developers, the model removes the transcription and captioning steps that multimodal agent pipelines usually require.

The release matters beyond Cloudflare's platform. Decision models — systems that return typed, probability-calibrated answers instead of generated text — are emerging as a cheaper, faster alternative to large language models for routing, classification and tool-selection work inside AI agents. Cloudflare is now shipping open-weight models in this category under Apache-2.0, while its closest competitor, TypeSafe AI's Jev, remains API-only in early access.

What is Clef-omni, exactly?

Clef-omni is a 30B total / 3B active parameter MoE post-trained from Qwen3-Omni-30B-A3B-Instruct. Cloudflare kept the comprehension backbone and dropped the text-to-speech output path. It runs on Workers AI as @cf/cloudflare/clef-omni with a 64,000-token hosted context, and Cloudflare says self-hosting was tested on a single H200, using about 64 GB of GPU memory in bfloat16.

Unlike a generative model, a decision model returns probabilities for every allowed option rather than free-form text. You send a state and a schema of questions. Clef supports three question types: noul (true or false), choice and score. The format follows the Jev and SystemOne API from TypeSafe AI.

Architecturally, the model runs a single prefill pass over the full payload. Video frames and audio sync into one sequence for joint processing. A small joint schema head then scores the options: each valid option first gathers evidence from the input, and field vectors cross-attend across the full context to compute confidence scores. Cloudflare froze the Qwen3 backbone and trained LoRA adapters, combining label-smoothed cross-entropy with Brier score calibration.

Supported inputs on Workers AI, per the developer docs:

• Images: up to 4, PNG/JPEG/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

Latency is the selling point. 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 it perform on benchmarks?

Clef-omni dominates intent classification. Cloudflare's Decision Index run shows it leading all four tested models with 94.8 macro-F1 on BANKING77 and 97.7 macro-F1 on CLINC150+OOS. It also tops Amazon ESCI at 57.8.

It trails elsewhere. When2Call accuracy is 63.3, versus 80.97 for Jev. Home appliances case-exact accuracy is 69.3, versus 97.73 for Clef-flash. On TypeSafe's workflow evals, Clef-omni scores below the text-focused Clef in all 5 reported metrics. The model's worst results: 2.6 accuracy on POP909-CL and 47.4 on GPQA Diamond versus Jev's 78.3. Benchmark scores come from Cloudflare's announcement and were not independently verified.

The pattern is consistent: one model can now decide across four modalities, but it pays for that breadth with weaker accuracy on hard reasoning.

How does the Clef family stack up against Jev?

Cloudflare's own comparison table, drawn from its Decision Index, positions the four models like this:

• Clef-omni: 30B total / 3B active MoE, Qwen3-Omni base, text + image + audio + video inputs, 64k hosted context, $0.15 per M input tokens, Apache-2.0
• Clef: 27B dense, Qwen3.8-27B base, text + image + video frames, 64k context, $0.24 per M tokens, Apache-2.0
• Clef-flash: 9B dense, Qwen3.5-9B base, text + image + video frames, 24k hosted context (256k self-hosted), $0.038 per M tokens, Apache-2.0
• Jev: parameters and base model not disclosed, text-only inputs per Cloudflare, context not disclosed by TypeSafe (Cloudflare cites 32k), $0.042 per M tokens, API only in early access

On headline evals, Clef-omni wins BANKING77 (94.8 vs Clef's 94.20, Clef-flash's 90.93 and Jev's 79.74) and CLINC150+OOS (97.7 vs 97.43, 66.77 and 89.27). Jev wins When2Call at 80.97 and BRIGHT at 47.52 nDCG@10. Clef leads invoice processing primary-action accuracy at 86.2.

How much faster is hosted Clef now?

The same announcement brought a serving upgrade, not new weights. Cloudflare moved hosted Clef to SGLang and tuned its stack. Median latency fell from 262 ms to 152 ms at about 800 tokens — a 1.7x gain. At about 3,400 tokens it fell from 616 ms to 305 ms, a 2.0x gain. At about 16,000 tokens it dropped from 2,721 ms to 1,635 ms. Clef support lands upstream in SGLang 0.5.22, so self-hosters get the same speedups.

Why this matters

The Clef family now covers three price and capability points: Clef-flash at $0.038 per M tokens for cheap high-volume text decisions, Clef at $0.24 for harder reasoning, and Clef-omni at $0.15 for multimodal payloads. Every model ships with open Apache-2.0 weights on Hugging Face, while Jev — currently the accuracy leader on agentic evals like When2Call — remains closed and early-access. That asymmetry gives developers a migration path off any single vendor.

The multimodal jump also collapses a real pipeline cost. An agent that previously transcribed audio, captioned video and then classified the text can now do all three in one scored call at 1.5-second latency for short clips. Clef-omni's weakness on GPQA Diamond and workflow evals shows the trade: breadth costs reasoning depth. Teams building multimodal agents now choose between Cloudflare's open, fast, four-modality option and TypeSafe's stronger but text-only Jev — and the next round of Decision Index runs will show whether Cloudflare closes the reasoning gap without giving up the speed.