Qwen3 235B A22B (Reasoning) is a large-scale Mixture-of-Experts (MoE) language model developed by Alibaba Cloud's Qwen team. It is specifically designed to handle complex, multi-step tasks such as advanced mathematics, software engineering, and scientific deduction through a dedicated thinking mode. This architecture allows the model to generate internal chains of thought before providing a final response, significantly improving its performance on logical reasoning benchmarks compared to standard instruction-tuned models.
The model utilizes a sparse architecture with 235 billion total parameters, of which 22 billion are activated per token during inference. This configuration employs 128 specialized experts, with 8 selected for each request to balance high-level intelligence with computational efficiency. It features 94 layers and utilizes Grouped Query Attention (GQA) and Rotary Positional Embeddings (RoPE) to optimize its 256K-token native context window, which is extendable up to 1 million tokens for long-form document analysis.
In terms of capabilities, the reasoning variant is optimized for high-precision outputs in technical domains. It has demonstrated state-of-the-art results for open-weight models on benchmarks like AIME 2025 and LiveCodeBench. Beyond its reasoning focus, the model maintains robust agentic features, natively supporting the Model Context Protocol (MCP) and complex tool-calling workflows, while providing multilingual support across more than 100 languages.