Qwen3 Omni 30B A3B (Reasoning) is a multimodal Mixture-of-Experts (MoE) language model developed by Alibaba's Qwen team. As the "thinking" variant of the Qwen3 series, it is specifically optimized for complex reasoning tasks, including advanced mathematics, coding, and logical deduction, by utilizing an integrated chain-of-thought (CoT) processing mechanism.
The model is natively omni-modal, designed to process text, image, audio, and video inputs end-to-end. This native integration allows for real-time streaming responses and improved cross-modal understanding compared to systems that rely on separate modality-specific encoders. The reasoning version is characterized by its ability to generate internal thinking tokens, which allow the model to verify its logic and refine its responses before producing final output.
Built on a Mixture-of-Experts architecture, the model consists of approximately 30.5 billion total parameters, with 3.3 billion parameters activated per token. This design aims to achieve the performance of much larger dense models while maintaining lower inference costs. It supports a standard context window of 256,000 tokens, which can be extended to handle long-form documents and high-complexity agentic workflows.