What is Cloudflare Clef-omni? Clef-omni is a newly released open-weight multimodal decision model that reads text, images, audio and video and scores answers to typed questions. Cloudflare announced the addition on October 9, 2026, extending its earlier Clef and Clef-flash family. Unlike a general chat model, Clef-omni produces probabilities over defined options, making it relevant to routing, classification and evidence-based decisions in applications rather than free-form conversational generation.
Cloudflare says the checkpoint is post-trained from Qwen3-Omni-30B-A3B-Instruct. The official model card calls it a 30B-A3B mixture-of-experts model and documents a joint schema head layered on a multimodal backbone. Inputs may combine JSON or text with images, speech recordings and video clips. One forward pass produces option scores for each question. This is a different output contract from ordinary text generation, so a benchmark for chat completion should not be treated as a direct comparison.
The weights are not merely promised. Cloudflare's Hugging Face repository lists sharded Safetensors for the backbone, joint_head.safetensors for the schema head, processor and tokenizer files, loader code and an Apache-2.0 license. The card describes tested execution using an NVIDIA H200 and approximately 64 GB of GPU memory for the backbone in BF16. These are publisher-provided deployment details, not an OMW hardware measurement or a guarantee about another device, quantization or runtime.
Cloudflare reports faster Clef-family inference and gives example multimodal latencies, including text decisions near 130 milliseconds at the median. Those claims depend on the provider's serving environment and workloads; Open Model Weights has not reproduced them. For teams evaluating self-hosted decision systems, the key questions are whether their schema maps cleanly to the head, whether media decoding meets privacy requirements and whether the model's exact revision, weight files and licensing are checked independently. Clef-omni is a new registry candidate, not automatically a field-verified entry.
What this changes for the evidence layer
New model intake candidate: verify Cloudflare/clef-omni at a pinned Hub revision, check the Qwen3-Omni base, separate backbone and joint schema-head Safetensors, capture Apache-2.0 licensing, and verify runtime dependencies. The decision head's scores are not free-form generated answers; latency and benchmark results are Cloudflare-reported, not OMW-measured.
Cloudflare official release and Hugging Face model card
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗