Cloudflare Launches Clef-omni with Full Multimodality
Cloudflare has expanded its open-weight decision model family with the introduction of Clef-omni, which natively processes audio, video, images, and text in a single pipeline.
Radwan AbdulhadiEditor and technology writer
Introduction of Clef-omni
Cloudflare has announced the expansion of its open-weight decision model family with the release of Clef-omni. Following the previous release of Clef and Clef-flash, the new model introduces native support for audio and video input alongside text and images, changing the paradigm for decision models which have traditionally been text-only.
Developers can explore the new capabilities and access technical details through the Cloudflare Blog announcement. The open weights are also available on HuggingFace for community use and deployment.
Multimodal Capabilities and Architecture
Clef-omni is built on a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation, allowing it to process text, imagery, audio, and video directly within a single pipeline. Rather than setting up cascading transcription pipelines or splitting audio and image channels from video, developers can call a single model to make decisions across any modality.
The model executes a quick prefill pass across the complete payload, scoring all modalities and valid parameter options simultaneously. By skipping output token generation, Clef-omni removes the overhead of transcribing or captioning incoming files, mapping media elements straight into a unified sequence.
Performance and Benchmarking
The architecture brings several core upgrades to the family, allowing calibrated, schema-bound decisions to seamlessly handle text, images, audio, and synchronized video in a single API call. Text-only decisions return in about 130 milliseconds at the median, while image inputs take about 150 milliseconds and audio clips take a few hundred milliseconds. A full 21-second video clip with sound is scored in approximately 1.5 seconds.
Cloudflare provides benchmark evaluations comparing the model against existing standards. Interested users can review TypeSafe evals to show how Clef-omni performs across various testing parameters.
Pricing Adjustments and Clef-flash Updates
Alongside the new model launch, Cloudflare has restructured pricing for its existing lineup. Clef-flash has been reduced from $0.09 per million input tokens down to $0.038 per million input tokens, making it cheaper than Jev while maintaining performance.
To achieve this lower price point, the hosted version of Clef-flash now features a reduced context window of 24k tokens, down from 64k. According to usage data, only 0.24% of requests exceed 24k input tokens. The model weights on Hugging Face remain untouched and retain support for a 256k context window for self-hosted deployments.
Serving Infrastructure and Speed Boosts
Cloudflare also boosted Clef inference speeds by up to 2.0x through infrastructure-level optimizations on Workers AI. These improvements were achieved primarily by migrating to SGLang for serving the model.
The engineering team collaborated on pull request #42721 to incorporate Clef into the SGLang project. Developers looking to self-host can find updated launch commands and technical guidelines in the official developer docs.
Sources
- Cloudflare BlogIntroducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash
Continue chronologically
Related entity coverage
Related stories
Cloudflare Remediates KV Cache Exposure in GLM-5.3
Cloudflare addresses a cross-request KV cache exposure issue in the GLM-5.3 Workers AI model through configuration updates and SGLang patches.
Cloudflare Releases Open-Source Clef Decision Models
Cloudflare introduces Clef and Clef-flash, open-source decision models hosted on Workers AI alongside a new reinforcement learning platform.
Cloudflare AI Search Launches Into General Availability
Cloudflare's AI Search reaches general availability, introducing native image embeddings, PDF OCR, larger file limits, and new billing terms.
Cloudflare ML uncovers evasive storefront attacks
Cloudflare says its Page Shield ML detected eight malicious JavaScript payloads that other scanning services failed to flag in live storefront traffic.
