October 12 — Cloudflare has released Clef-omni, a 30-billion-parameter decision model that scores text, images, audio and video in a single call instead of generating text. It launched on Workers AI on October 9, just eight days after the original Clef and Clef-flash. Why it matters: agent pipelines usually need separate transcription and captioning steps before anything can be scored — Clef-omni is built to skip them.
What is Clef-omni?
Clef-omni is the newest member of Cloudflare’s Clef family of decision models. A decision model does not write prose. You send it a state — a support ticket, a claim, a piece of content — plus a schema of typed questions, and it returns a calibrated probability for each allowed answer. The Clef format supports three question types (true/false, choice and score) and follows the System One API style first used by TypeSafe AI’s Jev model.
The new model is a mixture-of-experts system with 30 billion total parameters and about 3 billion active per call, post-trained from Alibaba’s Qwen3-Omni-30B-A3B-Instruct with the speech-output path removed. It is available on Workers AI as @cf/cloudflare/clef-omni, priced at $0.15 per million input tokens, with a 64,000-token hosted context window. The weights are published under the Apache 2.0 licence, and Cloudflare says self-hosting has been tested on a single H200 GPU using roughly 64 GB of memory.
What can it take in, and how fast is it?
Per Cloudflare’s developer documentation, one request can carry up to four images (PNG, JPEG or WebP), up to four audio clips (WAV or MP3, up to 300 seconds each), and up to two video clips (MP4 or WebM, up to 60 seconds each), alongside text or JSON — with between 1 and 64 questions per request. Cloudflare reports median decision times of about 130 milliseconds for text and about 150 milliseconds for images, with a 21-second video with sound scored in roughly 1.5 seconds. Those are the company’s own figures, not independent measurements.
How does Clef-omni perform against Clef and Jev?
Honestly: it depends on the task. In Cloudflare’s own Decision Index run, Clef-omni leads the family on intent classification, scoring 94.8 macro-F1 on BANKING77 and 97.7 on CLINC150+OOS — the best of the four models compared. It trails on harder reasoning and workflow evaluations, including 47.4 on GPQA Diamond against Jev’s 78.3, and 63.3 accuracy on When2Call against Jev’s 80.97. In other words, adding audio and video breadth cost it some depth on text-only reasoning. These benchmarks are vendor-run and have not been independently verified.
What else changed in the October 9 update?
Two things developers will feel immediately. First, Clef-flash got cheaper: its price fell from $0.09 to $0.038 per million input tokens — undercutting both TypeSafe’s Jev and Microsoft’s Decision-1, each at $0.042 — though its hosted context window narrowed to 24,000 tokens (the open weights still support 256,000 for self-hosting). Second, the original hosted Clef got faster without new weights: after a move to the SGLang serving stack, Cloudflare reports median latency improvements of roughly 1.7x to 2x depending on input size, at the same $0.24 per million tokens.
Why it matters
Decision models are quietly becoming one of the most contested layers in the AI stack: Cloudflare, TypeSafe AI and Microsoft have all shipped or repriced one in the last two weeks, and the fight is over the unglamorous work — routing, moderation, scoring, guardrails — that agents do thousands of times a day. Clef-omni’s bet is that one multimodal call beats a chain of specialist models. If its accuracy holds up on your workload, that is a real pipeline simplification. Everything AI, tested: the benchmark caveats above are exactly why you should run your own eval before switching.
Frequently asked questions
Is Clef-omni a chatbot?
No. It returns probabilities for a fixed set of questions you define, rather than generating free-form text — which makes its output easy to feed straight into routing, moderation or approval workflows.
How much does Clef-omni cost?
Cloudflare lists it at $0.15 per million input tokens on Workers AI. The smaller Clef-flash now costs $0.038 per million, and the original Clef remains $0.24.
Can I run Clef-omni myself?
Yes. The weights are on Hugging Face under Apache 2.0, and Cloudflare says it has tested self-hosting on a single H200 GPU in bfloat16, using about 64 GB of memory.
Sources: Cloudflare announcement and Workers AI documentation (October 9, 2026), via MarkTechPost and Tech Insider reporting