Qwen3-235B-A22B-Thinking-2507

Qwen3-235B-A22B-Thinking-2507

Qwen3-235B-A22B-Thinking-2507 is a state-of-the-art open-source Mixture of Experts (MoE) language model designed for deep reasoning tasks. With 235 billion parameters (22 billion active), it achieves exceptional performance across a wide range of reasoning-intensive applications, including logical reasoning, mathematics, science, coding, and academic benchmarks. The model demonstrates significant advancements in instruction following, tool usage, text generation, and alignment with human preferences, setting new standards for open-source "thinking" models.

This iteration of Qwen3 supports an extended 256K context length, enabling it to handle highly complex reasoning tasks that require long-context comprehension. It is optimized for "thinking mode," where outputs are structured to reflect deep reasoning processes. The model’s architecture features 94 layers, 128 experts (8 activated per query), and 64 attention heads for Q with 4 for KV, ensuring robust computational efficiency and scalability.

Qwen3-235B-A22B-Thinking-2507 is pre-trained and post-trained to achieve superior results on benchmarks such as MMLU-Pro, AIME25, HMMT25, and LiveCodeBench. It also excels in multilingual tasks and agentic applications, showcasing strong tool-calling capabilities via the Qwen-Agent framework. This allows seamless integration with external tools and APIs, simplifying development workflows.

Available on Hugging Face and via API, the model is supported by popular frameworks such as Transformers, SGLang, and vLLM. Its deployment options include OpenAI-compatible API endpoints, and it is compatible with local applications like Ollama and LMStudio. Recommended usage includes leveraging its extended token capacity (32,768 to 81,920 tokens) for complex queries and optimizing sampling parameters for improved performance.

Qwen3-235B-A22B-Thinking-2507 represents a cutting-edge solution for developers and researchers seeking advanced reasoning capabilities in open-source AI models.