Cloudflare Launches Clef-omni Multimodal Decision Model for Audio and Video

Cloudflare released Clef-omni on October 9, 2026, introducing a multimodal decision model capable of scoring text, images, audio, and video in a single API call. According to MarkTechPost, this open-weight addition to the Clef family arrives eight days after the initial launch of Clef and Clef-flash, expanding the vendor’s decision-making toolkit for developers.

The model is built on Qwen3-Omni-30B-A3B and operates with 30 billion total parameters, of which 3 billion are active. It runs on Workers AI and can be self-hosted on a single H200 GPU requiring about 64 GB of memory. Cloudflare priced Clef-omni at $0.15 per million input tokens, while simultaneously lowering Clef-flash costs to $0.038 per million tokens.

Technical Specifications and Performance

Clef-omni processes multimodal inputs by syncing video frames and audio into a single sequence for joint processing. The model supports up to four images, four audio clips up to 300 seconds each, and two video clips up to 60 seconds each. According to developer documentation, it handles a 64,000-token context window on Workers AI.

Performance testing shows Clef-omni achieves a 97.7 macro-F1 score on CLINC150+OOS, topping four Clef and Jev models. However, it lags behind text-focused counterparts on complex reasoning tasks, scoring 2.6 accuracy on POP909-CL and 47.4 on GPQA Diamond. Text decisions return in about 130 milliseconds, while a 21-second video with audio scores in approximately 1.5 seconds.

Impact on Development Workflows

This release aims to simplify multimodal agent pipelines by removing the need for separate transcription and captioning steps. Developers can send a state and a schema of questions to receive typed answers with probabilities. The model supports three question types: noul, choice, and score, following the API format established by TypeSafe AI’s Jev and SystemOne models.

Cloudflare froze the Qwen3 backbone and trained LoRA adapters for Clef-omni, using label-smoothed cross-entropy and Brier score calibration. The update also made hosted Clef roughly twice as fast. This architecture allows the model to gather evidence from inputs and cross-attend across the full context to compute confidence scores for each valid option.

What should you do?

Security and technology professionals should evaluate Clef-omni for applications requiring rapid, multi-modal classification without generating full text responses. Teams should review the model’s performance metrics, particularly its lower accuracy on hard reasoning benchmarks, to ensure it meets their specific reliability requirements.

What to do and how to stay safe: Cloudflare

• Review the Apache-2.0 license terms to confirm compliance with your organization's software usage policies before integrating the model.
• Benchmark the model’s latency and accuracy against your current transcription and captioning pipelines to assess efficiency gains.
• Monitor the model’s performance on complex reasoning tasks, as it may underperform compared to specialized text-only models.
• Wait for official documentation updates or vendor patches if you identify critical accuracy gaps in your specific use case.

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is the pricing for Cloudflare Clef-omni?

Clef-omni costs $0.15 per million input tokens, while Clef-flash was reduced to $0.038 per million tokens.

What inputs does Clef-omni support?

It supports up to four images, four audio clips, and two video clips, along with text, in a single call.

How does Clef-omni perform on reasoning tasks?

It scores lower than text-focused Clef on hard reasoning tasks, such as GPQA Diamond and POP909-CL.