What changed
Cloudflare released Clef-omni on October 9 as an open-weight multimodal decision model. It accepts text or JSON plus images, audio, or video, then answers a schema of typed questions—true/false (noul), named choice, or ordered score—with probabilities in one forward pass. The product is exposed through Workers AI as @cf/cloudflare/clef-omni; Cloudflare also published Apache-2.0 weights.
Why it matters
Most multimodal agent pipelines are a chain: transcribe, OCR, prompt a chat model, parse free-form output, then guess confidence. Clef-omni makes the decision contract explicit. A support-triage tool can ask for urgency, team, and severity in one request; an inspection workflow can combine a photo, audio clip, and video without separate modality-specific models. Typed answers and probabilities make routing, thresholds, and audits easier to test.
Product takeaway
This is a useful split between “generate an explanation” and “make a bounded decision.” Use a general model for narratives; use a decision model when downstream code needs a small, inspectable state machine. Cloudflare’s model documentation lists a 64,000-token context window and $0.15 per million input tokens with no output-token charge, which makes the hosted path interesting for high-volume classifiers and routers.
The tradeoff is equally clear: self-hosting is not lightweight. The open-weight model card says the release was tested on a single H200 and needs about 64 GB of BF16 GPU memory. For most teams, the practical starting point is the hosted API; the weights keep a portability and inspection path open for teams that need it.