Cloudflare Clef-omni: A Decision Model for Audio and Video

Cloudflare Clef-omni is a new open-weight decision model that can look and listen as well as read. Announced on the Cloudflare blog on October 9, 2026, Clef-omni accepts text, images, audio and video in a single API call and returns a probability for each allowed answer. It joins Clef and Clef-flash, the two decision models Cloudflare introduced on October 1, and it arrives with a price cut and speed gains for the existing pair.

• What it is: a multimodal decision model that takes text, images, WAV or MP3 audio, and MP4 or WebM video in one request.
• Price: $0.15 per million input tokens on Workers AI.
• Speed: about 130 ms for text and about 1.5 seconds for a 21-second video with sound, per Cloudflare.
• Open weights: published on Hugging Face under Apache 2.0, built on Qwen3-Omni-30B-A3B-Instruct.

What Is a Multimodal Decision Model?

Decision models are a young category. In place of generating text, they score a fixed set of options and return calibrated probabilities, which makes them a fit for routing, moderation and tool selection inside agents. The Hugging Face card for Clef-omni puts it plainly: "There is no free-form text generation and no output parsing."

What Clef-omni adds is input range. Earlier models in the category, including Jev from TypeSafe AI and Microsoft-Decision-1, work from text or structured state. Clef-omni takes audio and video directly, with frames and sound kept in sync, so a developer does not need a separate speech-to-text or frame-extraction step before asking a question such as whether a clip breaks a policy.

Cloudflare built it by post-training Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts model, and removing the speech-output components that a decision model has no use for.

How Fast and How Much?

Cloudflare reports median latency of about 130 ms for text-only decisions and about 150 ms with an image. "Even a full 21-second video clip with sound is scored in about 1.5 seconds, all in a single API call," the company writes.

The same post updates the rest of the family:

• Clef-flash: cut from $0.09 to $0.038 per million input tokens.
• Clef: unchanged at $0.24, but faster through serving changes alone, with median latency on inputs of about 3,400 tokens falling from 616 ms to 305 ms.
• Clef-omni: new at $0.15.

One trade-off is stated openly. The hosted Clef-flash context window drops from 64k to 24k tokens; Cloudflare says only 0.24% of requests exceeded that size, and the open weights still support the longer context for anyone self-hosting.

How Does Clef-omni Compare With Jev?

The comparison figures come from Cloudflare's own testing, so they should be read as vendor results. On that basis the picture is mixed in an informative way. Cloudflare reports Clef-omni ahead of Jev on intent classification, with 94.8 against 79.74 macro-F1 on BANKING77, and on the BFCL function-calling test at 98.2 against 95.75. Jev leads on When2Call, 80.97 to 63.3, and on the BRIGHT retrieval benchmark.

The model card goes further and shows Jev well ahead on harder reasoning tests such as GPQA Diamond, at 78.3 against 47.4. I appreciate that the weaker results are published alongside the wins. They suggest a sensible division of labor: Clef-omni for fast perceptual and classification calls, and a stronger reasoner where the decision is hard.

Why It Matters

Cloudflare says it already uses Clef internally to close spam issues on its documentation repository, to moderate plugins for phishing, to scan for personal data and to help flag malicious domains. Those are high-volume jobs where a few hundred milliseconds and a fraction of a cent per call change what is practical.

With Apache 2.0 weights that Cloudflare tested on a single H200 GPU, teams can also run the model themselves. An open decision model that handles sound and video widens what this new class of system can be used for. More in our AI coverage.

Sources: Cloudflare Blog — Clef: faster, cheaper, multimodal — October 9, 2026; Hugging Face — Cloudflare/clef-omni model card — October 2026.