The Pulse
Cloudflare's Clef-Omni Takes Audio and Video in One Call
Cloudflare released Clef-omni, an open-weight decision model that accepts audio, video, images and text. The company says it can score a 21-second video with sound in about 1.5 seconds, while also cutting Clef-flash’s hosted price and impro
AI.info Team · Written with AI ·
Cloudflare says its new decision model can score a 21-second video clip with sound in about 1.5 seconds. Released on October 9, Clef-omni accepts audio, video, images and text in one request, then returns structured decisions rather than a passage of generated text.
The release adds audio and native video input to Cloudflare’s Clef family, whose earlier models already handled text, images and video-frame arrays. The company says developers can combine evidence from a photo, a recording and a video in one call instead of routing each through separate transcription or vision systems.
One decision from sound, images and video
Cloudflare’s example asks whether an appliance’s model label is visible in a photo, whether its operating sound is normal, and whether its fan is running in a video. Clef-omni scores answers to those typed questions from the combined input. It supports audio files in WAV or MP3 formats and video in MP4 or WebM, alongside text and images.
“Even a full 21-second video clip with sound is scored in about 1.5 seconds, all in a single API call,” write Michelle Chen, Alex Reneau and Kevin Jain, the authors of Cloudflare’s announcement.
Built on Qwen3-Omni, without speech generation
Clef-omni uses the comprehension backbone of Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts model, while dropping its text-to-speech components, Cloudflare says. Instead of generating output tokens, the model processes the combined media and scores valid answers to the questions supplied by the developer.
Cloudflare describes a two-stage attention system that gathers evidence for each answer option from text, image, audio and video input, then calculates confidence scores across the full context. The company says its training keeps the Qwen3 backbone frozen, uses low-rank adapters and combines label-smoothed cross-entropy loss with Brier-score calibration.
Benchmark results vary by task
Cloudflare’s published results do not show Clef-omni leading every comparison. On BFCL case exact, it scores 98.2, below Clef at 98.47 and Clef-flash at 98.76, but above TypeSafe’s Jev at 95.75. On the When2Call accuracy test, Clef-omni scores 63.3, compared with Jev’s 80.97.
For other reported measures, Clef-omni scores 92.7 on API-Bank accuracy and 97.7 on CLINC150+OOS macro-F1. The figures are Cloudflare’s own benchmark results; the announcement does not describe an independent evaluation. Its TypeSafe workflow comparisons also show mixed results, with Clef-omni below Jev on invoice-processing exact actions and agent-trace observability’s primary-action measure.
Cheaper Clef-flash comes with a shorter hosted context
Cloudflare also cut Clef-flash’s hosted input price from $0.09 to $0.038 per million tokens. Clef-omni launches at $0.15 per million input tokens, while Clef remains at $0.24. The lower Clef-flash price comes with a hosted context window reduced from 64,000 tokens to 24,000; the company says the downloadable weights remain unchanged and support a 256,000-token context when self-hosted.
Cloudflare reports that only 0.24% of its Clef-flash requests exceed 24,000 input tokens. It also reports serving-speed improvements for hosted Clef: median latency fell from 262 to 152 milliseconds for inputs of about 800 tokens, and from 616 to 305 milliseconds for inputs of about 3,400 tokens. Those are infrastructure optimizations, the company says, so it released no new Clef weights with the speed update.