Cloudflare adds audio and video input to Clef decision models

Cloudflare has launched Clef-omni, a decision model that can analyze audio, video, images, and text in one request. Michelle Chen, Alex Reneau, and Kevin Jain announce the update for the Cloudflare Blog, alongside faster performance for Clef and a major price reduction for Clef-flash.

Decision models are designed to return structured choices, such as labels, classifications, or actions, rather than generate long passages of text. We have a separate article explaining them. Cloudflare says Clef-omni can process media without requiring users to build separate steps for transcription, captioning, or extracting video frames.

For example, a business could submit a product image, an audio recording of a machine, and a short video, then ask whether a label is visible, whether the equipment sounds normal, and whether a fan is operating. The model returns answers for the predefined fields.

Faster multimodal decisions

Clef-omni is based on Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts model. Cloudflare says its version uses the underlying model to interpret several media formats but does not generate speech or lengthy text. According to the company, median response times are about 130 milliseconds for text, 150 milliseconds for images, several hundred milliseconds for audio, and roughly 1.5 seconds for a 21-second video with sound.

Cloudflare has set Clef-omni pricing at $0.15 per million input tokens. Clef-flash now costs $0.038 per million input tokens, down from $0.09, while the standard Clef model remains at $0.24.

The lower Clef-flash price comes with a smaller hosted context window of 24,000 tokens, reduced from 64,000. Cloudflare says only 0.24 percent of its requests exceed that limit. Self-hosted Clef-flash weights still support up to 256,000 tokens, according to the company.

Cloudflare also reports that infrastructure changes, including a move to SGLang for serving, make hosted Clef up to twice as fast for some input sizes. The Clef family is available through Workers AI and is compatible with the Jev API, Cloudflare says.