#

Key takeaways

• Clef-omni accepts text, images, audio, and video in one request.

• It scores caller-defined answer options instead of generating free-form text.

• Cloudflare released open weights on Hugging Face and offers the model on Workers AI.

• Clef-flash now has a lower hosted input price and a smaller hosted context window.

• Cloudflare says hosted Clef serving is faster after infrastructure changes.

What Cloudflare announced

Cloudflare's October 9 announcement adds a new model to the Clef family: Clef-omni. The company describes it as a decision model for multimodal inputs. It can take text, images, audio, and video together in one request.

Cloudflare says the model does not generate free-form text. Instead, it scores answer options that the caller defines. That makes the model fit structured decision tasks rather than open-ended generation.

The company says Clef-omni is available on Workers AI. It also released open weights on Hugging Face. The announcement says the implementation uses a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts backbone and removes speech-output components.

How the model is meant to work

The source material includes a documented API example. In that example, a request includes an image, an audio recording, and a video. The request also includes structured questions about an appliance.

That example shows the model's intended pattern. It combines multiple media types and asks the model to choose among defined answers. This is a practical setup for decision workflows where the output needs to stay structured.

Clef-omni joins the Clef family introduced the previous week. The October 9 update adds native audio and video input alongside text and images.

Updates to Clef and Clef-flash

The same October 9 post also reports changes to other Clef offerings. Cloudflare says hosted Clef-flash input pricing drops from $0.09 to $0.038 per million input tokens. It also says the hosted context window changes from 64,000 tokens to 24,000 tokens.

Cloudflare adds an important distinction. It says the released weights are unchanged and still retain a longer trained context capacity for self-hosting. That means the hosted service and the self-hosted weights are not described in the same way.

The post also says the original Clef model's hosted serving is now up to twice as fast in Cloudflare's measurements. Cloudflare attributes that to infrastructure changes. It says the model weights themselves are unchanged.

Benchmarks and interpretation

Cloudflare lists benchmark results for Clef-omni, Clef, Clef-flash, and Jev across decision tasks. These are vendor-reported figures. They vary by task.

That matters for interpretation. The source does not present the benchmarks as independent proof of universal superiority. It also does not claim that one model is best for every workload.

The safest reading is narrower. The announcement shows that Cloudflare is positioning these models for decision tasks, with different tradeoffs in modality, speed, pricing, and deployment style.

Operational considerations

The source points to a few practical considerations. First, hosted pricing and context limits can change. Second, latency may vary by deployment. Third, availability may also vary by deployment and future changes.

The announcement also separates hosted behavior from released weights. That distinction matters for teams that want to self-host. The source says the released weights keep a longer trained context capacity, while the hosted Clef-flash context window is smaller.

Because the model scores caller-defined options, users need structured prompts and answer sets. That is not a risk in itself, but it is an operational requirement. The model's output depends on how the options are framed.

Morocco relevance

The source reports no Morocco-specific launch, customer, regulation, or measured effect. A possible global lesson is that multimodal decision models may fit structured workflows where outputs must stay constrained.

Bottom line

Cloudflare's update expands the Clef family with a multimodal decision model and adjusts pricing and serving details for existing models. The announcement is strongest on product positioning and deployment options.

It is also careful about what it claims. The benchmarks are vendor-reported, and the source does not present them as universal proof. For readers, the main takeaway is simple: Cloudflare is pushing structured, multimodal decisioning with different hosted and self-hosted tradeoffs.