Cloudflare expands Clef with multimodal input, lower costs and faster responses

Cloudflare has bundled three changes to its Clef decision-model family rather than presenting a single isolated model release. The company introduced Clef-omni, described as the first Clef model able to handle audio, video, images and text. At the same time, Clef-flash has become cheaper, while the standard Clef has been made faster. Cloudflare’s own shorthand for the update is “faster, cheaper, multimodal.”

What changed

• Clef-omni broadens the input surface. The model is designed to work across four modalities: audio, video, images and text. That gives decision systems more ways to gather evidence and interact with users. However, the available material does not include benchmark scores, context limits, supported formats or task-level accuracy. Multimodal support alone should therefore not be treated as proof of overall superiority.
• Clef-flash focuses on cost. Cloudflare says the model is now cheaper, but the supplied information does not state a price or a percentage reduction. The practical objective is clear: make repeated calls more economical for automation, high-volume workflows and applications that need frequent decisions.
• Clef focuses on speed. The company says the model is faster, although no latency, throughput figures or testing methodology are provided. The safest interpretation is that Cloudflare is improving the serving experience, not that the model has been definitively ranked ahead of alternatives.
• Clef-omni builds on an open model. The new model is based on Qwen3-Omni-30B-A3B-Instruct, an open-source model identified in the supplied summary as using a mixture-of-experts design. The available information does not explain what Cloudflare changed in training, alignment, deployment or inference.

Why it matters

The release reflects a more practical phase of competition around decision models. Earlier discussions often centered on whether a model could reason through a difficult prompt. Once such systems are embedded in workflows, agents or real-time products, other constraints become just as important: can the system accept the necessary inputs, can each call be afforded, and can it respond quickly enough?

Clef-omni could expand the kinds of evidence a decision system can use. A model that sees an image, listens to audio or processes video is not limited to text-based instructions and records. Yet multimodal operation also introduces additional compute and engineering complexity. Its real value will depend on reliability, error rates, supported use cases and total serving cost—details that are not available in the current material.

Taken together, the three updates look like a product-oriented upgrade rather than a single capability announcement. Clef-omni addresses input diversity, Clef-flash addresses economics, and Clef addresses responsiveness. The information does not establish a decisive technical lead, but it does show where the market is moving: decision models are increasingly being evaluated on capability, price and speed at the same time.

Source: OSChina