A week after Clef, Cloudflare has launched Clef-omni, which can make decisions based on text, images, audio and video in a single call, and has reduced the price of Clef-flash.
Cloudflare continues. A week after launching Clef and Clef-flash, its first open-weight "decision models", the company unveils Clef-omni, which accepts audio (WAV, MP3) and video (MP4, WebM) as inputs in addition to text and images. It also announces a reduction in the price of Clef-flash, now cheaper than Jev, the TypeSafe model that popularized the genre, and an acceleration of Clef.
No text generation
A decision model does not write a response: it receives a status and a list of constrained choice questions, then returns a value for each field accompanied by a confidence score. Clef-omni is based on Qwen3-Omni-30B-A3B, a model with a mix of experts whose understanding Cloudflare keeps the core of understanding by discarding speech synthesis, and which it adapts by LoRA with a Brier score calibration. In concrete terms, the same call can check on a photo that a label is legible, by ear that a device is running without abnormal noise, and on a video that a fan is working, without an intermediate transcription chain.
In terms of performance, Cloudflare announces a median latency of about 130 ms for a text decision, 150 ms with an image and a few hundred milliseconds for an audio clip. The model is available on Workers AI and its weights are published on Hugging Face.
Read the full article on the original source
Go to the source →