Qwen3 235B A22B Thinking 2507

Qwen3-235B-A22B-Thinking-2507 is a Mixture-of-Experts model built on 94 transformer layers with grouped query attention, sparse gating across 128 experts where only 8 activate per token, and 234B non-embedding parameters. This architecture keeps computational costs manageable while enabling deep, multi-step reasoning across complex domains. The model is purpose-built as a thinking-only variant, enforcing structured reasoning through a special </think mode rather than serving as a general-purpose chatbot. It is designed for highly complex tasks that benefit from extended deliberation, supporting long-form generation in mathematics, science, and programming with outputs reaching up to 81,920 tokens.

The model progressed through both pretraining and post-training stages, with recent refinements over three months that sharpened reasoning quality and depth across logical, mathematical, and coding benchmarks. Instruction tuning during post-training strengthened instruction following, tool usage, alignment with human preferences, and agentic workflows. Benchmark performance on AIME, SuperGPQA, LiveCodeBench, and MMLU-Redux places it at the leading edge among open-source thinking models, outpacing many closed alternatives in structured reasoning scenarios. It natively supports parallel function calling, response schemas, and multilingual tasks, making it well-suited for automation pipelines and research workflows. Released under an Apache 2.0 license, the model offers full weight access for enterprises and developers seeking to deploy state-of-the-art open reasoning without proprietary constraints.