On October 9, Cloudflare released Clef-omni, adding audio and full video input to its family of “decision models.” The weights are on Hugging Face under an Apache 2.0 license, and the hosted version runs on Workers AI. It arrives just eight days after the debut of Clef and Clef-flash.
Decision models do a deliberately narrow job. The caller supplies a state and a set of predefined questions, each with a few answer options, and the model assigns a probability to every option. It generates no free-form text, so there is no output to parse. The internal use cases Cloudflare cites include spotting and closing spam issues in GitHub documentation repositories, screening a plugin registry for phishing, scanning data for personal information such as national ID numbers, and flagging malicious domains.
Four kinds of input in one call
The earlier Clef handled text, images and frames extracted from video. Clef-omni takes WAV and MP3 audio and MP4 and WebM video directly, processing the soundtrack and the picture in sync. It still does no transcription or captioning: it runs a single prefill pass and scores all modalities and all options together.
Cloudflare’s published median latencies are about 130 milliseconds for text only, about 150 milliseconds for an image, a few hundred milliseconds for audio, and about 1.5 seconds for a 21-second video with sound.
The base is Alibaba’s Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts model. Cloudflare kept the understanding components, dropped the speech-synthesis parts, froze the backbone, fine-tuned it with LoRA, and added a small Transformer scoring head. According to the model card, it runs in bfloat16 on a single H200, and the backbone needs roughly 64GB of GPU memory.
Scores don’t beat the predecessor across the board
In Cloudflare’s own comparison table, Clef-omni trails the vision-only Clef on most items. On the ToolRet tool-retrieval benchmark it scores 66.6 against 69.19; on When2Call, which tests when to invoke a tool, 63.3 against 72.37; on the PhishNChips phishing benchmark, 73.2 against 79.60. The biggest gap is the home-appliance command task, 69.3 against 82.95. It comes out slightly ahead of Clef on BANKING77 (banking intent classification) and CLINC150+OOS (intent detection).
Measured against TypeSafe’s decision model Jev, Clef-omni trails by more than 17 points on When2Call and leads by 15 points on BANKING77. All of these scores are self-reported by Cloudflare, and the blog post does not go into evaluation details.
Two adjustments on the same day
The hosted price of Clef-flash fell from $0.09 to $0.038 per million input tokens, which Cloudflare says now undercuts Jev. The trade-off is that the hosted context window shrank from 64k to 24k, a change Cloudflare justifies by noting that only 0.24% of requests exceed 24k. The open weights on Hugging Face are unaffected and were still trained at 256k. Clef’s price is unchanged; after moving inference to SGLang, Cloudflare says median latency improved by 1.7 to 2.0 times.
Another Qwen foundation
Three days earlier, Microsoft’s Decision-1, another decision model built for multiple-choice judgments, was released on a Qwen 9B base. Within a single week, two US companies have shipped similar products, both built on Alibaba’s open models. For developers in China, the Clef weights are freely downloadable under a permissive license, and the API is declared compatible with Jev, so the barrier to building an in-house moderation, routing or ticket-triage service is fairly low. Cloudflare has not said anything specific about how reachable the hosted Workers AI service is from within China, so teams should test it themselves before using it in production.
Sources: Cloudflare official blog, Hugging Face model card (license, base model and GPU memory requirements), CocoLoop, IT Home, RuntimeWire; benchmark scores and prices are as self-reported by Cloudflare.