Overview
Clef-omni is Cloudflare open-weight multimodal decision model, released on October 9, 2026. Instead of generating text, it reads text, images, audio and video from a single request and returns a calibrated probability for every option the developer supplies, for structured judgements such as routing, classification and tool selection.
Key Features
- Four modalities in one call: A request can carry text, images, audio and video together, and the model scores them in a single prefill pass, removing the intermediate transcription and captioning steps
- Three structured question types: Supports true/false, multiple choice and ordinal score questions, up to 64 per request, returning a probability per option rather than free-form text
- Open weights for self-hosting: Weights ship under Apache-2.0; Cloudflare validated self-hosting on a single H200 in bfloat16 at roughly 64 GB of GPU memory
- Familiar decision interface: The request format matches existing decision models, so switching the model identifier is enough to try it
Use Cases
- Routing support tickets by intent and urgency, using the call recording and screenshots together as evidence
- Choosing between tools inside an agent, with probability thresholds deciding which API to call or whether to hand off to a human
- Equipment inspection from one photo plus a running-recording clip, judging whether a label is visible and whether the sound is abnormal
- Grading mixed image and short-video material in content moderation, with scores that drop straight into the review workflow
Pros
- Folding transcription and frame description into one call cuts both latency and information loss in multimodal pipelines
- Returns probabilities rather than text, so there is no parsing step before the result reaches routing logic
- Apache-2.0 open weights, usable both as a hosted service and on your own infrastructure
- Output is not billed separately, so cost tracks input tokens only
Pricing
On Cloudflare Workers AI it is billed at $0.15 per million input tokens with no output charge. Weights are released under Apache-2.0 for self-hosting, with a reference configuration of a single H200 in bfloat16 at about 64 GB of GPU memory, and a hosted context length of 64,000 tokens.
Summary
Clef-omni suits developers who need structured judgements inside an agent pipeline, especially when inputs mix visuals and sound and the desired output is a probability rather than a paragraph. Open weights make it easier to adopt for teams that self-host or keep data inside their own network.
Version History
- Cloudflare releases Clef-omni multimodal decision model (2026-10-09): Cloudflare released Clef-omni, the third member of the Clef family, post-trained from Qwen3-Omni-30B-A3B-Instruct with the backbone frozen and LoRA adapters plus a joint schema head trained on top. One call handles text, images, audio and video, accepting up to four images, four audio clips and two video clips with up to 64 questions per request. Weights ship under Apache-2.0 with a 64K hosted context; median latency is about 130 ms for text-only decisions, about 150 ms with images and about 1.5 seconds for a 21-second video with sound.