Qwen has released the Qwen3-235B-A22B-Thinking-2507 model, an update to its 235B parameter mixture-of-experts architecture designed for complex reasoning tasks. This version features significantly enhanced performance on logical reasoning, mathematics, science, and coding benchmarks, achieving state-of-the-art results among open-source thinking models.
The model maintains a total of 235 billion parameters with 22 billion activated per token, supporting a native context length of 262,144 tokens. It is optimized for instruction following, tool usage, and alignment with human preferences, while enforcing a dedicated thinking mode via default chat templates.
Qwen recommends this version specifically for highly complex reasoning tasks due to its increased thinking length and improved general capabilities.