Editorial illustration for Cloudflare Clef-omni: 30B-A3B multimodal model for audio, video, image and text
Cloudflare's Clef-omni: Multimodal AI for Audio, Video, Text
Cloudflare put a new model out eight days after its last one. Clef-omni landed October 9, 2026, following the October 1 launch of Clef and Clef-flash, and it's built to handle something those two couldn't: scoring text, images, audio and video in a single call. That's the pitch, anyway. No separate transcription pass, no captioning step bolted onto the front of a pipeline, just one model that takes in whatever modality you throw at it and returns an answer.
The model runs 30 billion total parameters with 3 billion active, a mixture-of-experts setup, and supports a 64,000-token context window on Workers AI under the name @cf/cloudflare/clef-omni. Cloudflare says it's also been tested running on a single H200 GPU for self-hosting, pulling around 64 GB of memory in bfloat16. That same October 9 update dropped pricing on Clef-flash below TypeSafe's Jev and roughly doubled the speed of hosted Clef.
Clef-omni isn't a clean upgrade across the board. It holds its own on some tasks and falls noticeably short on others, which raises the question of what a decision model actually is and where Clef-omni fits against its siblings.
Cloudflare has released Clef-omni, a new open-weight member of its Clef decision model family. Clef-omni is a Cloudflare decision model that scores text, images, audio and video in 1 call.
Why this matters For anyone building multimodal agents, Clef-omni's pitch is practical rather than flashy: one scoring call instead of a transcription-then-caption-then-score chain. That's a real engineering win if the quality holds up, since every extra pipeline stage is another place for latency and errors to creep in. Cloudflare shipping this just eight days after the original Clef, alongside a Clef-flash price cut undercutting TypeSafe's Jev, signals it's treating decision models as a competitive battleground, not a side project.
Still, the Brier-score calibration claim deserves scrutiny before anyone wires Clef-omni into production judging or moderation. Label-smoothed training helps with overconfidence, but calibration numbers from a vendor's own docs aren't the same as third-party eval results. The input caps matter too: four images, four 300-second audio clips, two video clips per call.
That's generous for most agent use cases, but teams scoring longer-form content will hit the ceiling fast. Worth testing against your own data before trusting the scores, not just the speed claims.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives New Jev Decision Index Ranks Open Reproductions at 64.1 Score
From Archives Execs Say Voice AI Hasn’t Reached Its ChatGPT Moment Yet, Point to Next Stage
Common Questions Answered
What makes Cloudflare Clef-omni different from the earlier Clef and Clef-flash models?
Clef-omni is a multimodal model that can process text, images, audio, and video in a single call, whereas the earlier Clef and Clef-flash models could not handle this combination of modalities. This eliminates the need for separate transcription passes or captioning steps that would normally be required in a traditional pipeline.
How many parameters does the Clef-omni model contain?
Clef-omni is a 30 billion parameter model, making it a substantial decision model designed to handle complex multimodal scoring tasks. The model size reflects Cloudflare's commitment to delivering quality performance across multiple input modalities.
What is the practical engineering advantage of Clef-omni's single-call approach?
By consolidating audio, video, image, and text scoring into one model call instead of a multi-stage pipeline, Clef-omni reduces latency and eliminates multiple points where errors could accumulate. This streamlined approach provides a real engineering win for developers building multimodal agents who would otherwise need to chain together transcription, captioning, and scoring steps.
When was Clef-omni released relative to other Clef models?
Clef-omni was released on October 9, 2026, just eight days after Cloudflare launched the original Clef and Clef-flash models on October 1, 2026. This rapid succession of releases demonstrates Cloudflare's aggressive push into the decision model market.
Further Reading
- Introducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash - Cloudflare Blog
- Clef-omni adds audio and video input, Clef-flash is now cheaper, and Clef is faster - Cloudflare Developers
- Cloudflare/clef-omni - Hugging Face
- Clef-omni — Cloudflare Workers AI model documentation - Cloudflare Developers
