The Qwen3-30B-A3B model has received a major update, introducing the new version Qwen3-30B-A3B-Thinking-2507. This new version has achieved significant improvements in reasoning ability, general capabilities, and context length, marking that the model is not only lighter but also has stronger thinking capabilities.
Compared to the Qwen3-30-A3B model released on April 29, the new model has made progress in several key areas. In mathematical ability, Qwen3-30B-A3B-Thinking-2507 scored 85.0 in the AIME25 evaluation, which exceeds Gemini2.5-Flash (thinking) and Qwen3-235B-A22B (thinking). In the code ability test LiveCodeBench v6, the new model also achieved a high score of 66.0, demonstrating its strong capability in programming. Additionally, the new model has shown significant improvement in knowledge level (GPQA, MMLU-Pro) compared to the previous version.
In addition to enhanced mathematical and coding abilities, Qwen3-30B-A3B-Thinking-2507 has achieved a comprehensive upgrade in general capabilities. In evaluations of writing (WritingBench), Agent capabilities (BFCL-v3), multi-turn dialogue, and multi-language instruction following (MultiIF), the new model outperformed Gemini2.5-Flash (thinking) and Qwen3-235B-A22B (thinking), showing its excellent ability in understanding and generating natural language.
The new model's context understanding capability has also been significantly enhanced, with native support for 256K tokens and the ability to scale up to 1M tokens. This improvement allows Qwen3-30B-A3B-Thinking-2507 to handle longer text input, thereby better understanding and generating complex text content. Additionally, the thinking length of the new model has been increased, and the development team recommends setting a longer thinking budget for highly complex reasoning tasks to fully leverage its potential.
To facilitate developers and researchers in using and testing this new version, Qwen3-30B-A3B-Thinking-2507 is now open-sourced on the ModelScope community and HuggingFace. Its lightweight size allows the model to be easily deployed locally on consumer-grade hardware. At the same time, the development team has also launched the new model on Qwen Chat, and we welcome developers to experience its powerful reasoning and generation capabilities.