Qwen3 235B A22B Thinking (2507)

Qwen3-235B-A22B-Thinking-2507 is the reasoning-specialized variant released in July 2025, activating 22 billion of 235 billion parameters per forward pass through mixture-of-experts routing. This thinking-only model is built for step-by-step problem solving in mathematics, science, coding, and complex logical reasoning, generating internal chains of thought in think blocks. It posts strong results among open-weight thinking models on benchmarks including AIME25 (92.3) and LiveCodeBench (74.1), and handles reasoning-heavy tasks that benefit from extended computation. It supports a 131K context window and strong multilingual reasoning across 100+ languages. The MoE design keeps computation efficient, activating only 22 billion parameters per token despite the 235 billion total, which makes it practical for demanding reasoning workloads. Open-weight under Apache 2.0.