Cloudflare Releases Clef-omni Multimodal Decision Model

Cloudflare released Clef-omni on October 9, 2026, introducing a new open-weight member to its Clef decision model family. This multimodal model is designed to score text, images, audio, and video inputs within a single API call. The launch occurred eight days after the initial release of Clef and Clef-flash. This update also included a reduction in Clef-flash pricing, positioning it below TypeSafe's Jev, and made hosted Clef approximately twice as fast. For developers, Clef-omni eliminates the need for separate transcription and captioning steps typically required in multimodal agent pipelines.

Clef-omni features a total of 30 billion parameters, with 3 billion active parameters utilizing a Mixture of Experts (MoE) architecture. It supports a substantial context window of 64,000 tokens and runs on Cloudflare's Workers AI platform, accessible via the identifier @cf/cloudflare/clef-omni. Self-hosting has been tested on a single NVIDIA H200 GPU, requiring approximately 64 GB of GPU memory when operating in bfloat16 precision. In terms of performance, Clef-omni demonstrates capabilities close to the original Clef model on routing and intent identification tasks. However, it performs less effectively on most reasoning and workflow evaluations compared to its predecessor.

One of Clef-omni's strengths is its high performance on specific benchmarks, achieving a macro-F1 score of 97.7% on the CLINC150+OOS dataset. This places it at the top among the four Clef and Jev models evaluated. Conversely, its weakest performance was observed on the POP909-CL benchmark, where it achieved an accuracy of 2.6%, and on the GPQA Diamond benchmark with 47.4% accuracy. For comparison, Jev achieved 78.3% accuracy on the GPQA Diamond benchmark. The primary advantage of Clef-omni is its ability to make a decision across four modalities using a single call, whereas its accuracy on complex reasoning tasks is noted as weaker than text-focused models.

A decision model, as defined by Cloudflare, returns typed answers with associated probabilities rather than generating free-form text. Users provide a state and a schema of questions, and the model outputs a probability for each allowed option. Clef supports three question types: noul (true or false), choice, and score, with its output format aligning with the Jev and SystemOne API specifications from TypeSafe AI. Clef-omni is built upon the Qwen3-Omni-30B-A3B-Instruct model. Cloudflare retained the comprehension backbone of Qwen3-Omni-30B-A3B-Instruct while removing its text-to-speech output capabilities. The model processes inputs through a single prefill pass, synchronizing video frames and audio into a unified sequence for joint analysis. A small joint schema head then assigns confidence scores to potential options. Each valid option is supported by evidence gathered from the input, with field vectors cross-attending across the entire context to compute these scores. Cloudflare employed LoRA adapters for training, freezing the Qwen3 backbone during this process. The training methodology combined these adapters with the frozen backbone.