Ali (09988) Tongyi Qianwen introduces new model Qwen3-30B-A3B-Thinking-2507.

On July 31st, Alibaba (09988) introduced a new model Qwen3-30B-A3B-Thinking-2507 to everyone, a more intelligent, agile, and versatile new reasoning model. Currently, Qwen3-30B-A3B-Thinking-2507 has been open-sourced on the MaDa Community and HuggingFace. Compared to the Qwen3-30-A3B model that was open-sourced on April 29th, the new model has shown significant improvements in reasoning ability, general ability, and context length: Mathematically strong, proficient in coding: The new model achieved a high score of 85.0 in the AIME25 evaluation focusing on mathematical ability and a score of 66.0 in the LiveCodeBench v6 coding ability test, surpassing Gemini2.5-Flash(thinking) and Qwen3-235B-A22B(thinking) in both core reasoning abilities. The knowledge level (GPQA, MMLU-Pro) of the new model has also significantly improved compared to the previous version. Intuitive understanding, comprehensive upgrade: In general ability evaluations such as WritingBench, Agent capability (BFCL-v3), multi-turn conversation, and multi-language instruction compliance (MultiIF), Qwen3-30B-A3B-Thinking-2507 outperforms Gemini2.5-Flash(thinking) and Qwen3-235B-A22B(thinking). Greater context understanding: Native support for 256K tokens, expandable to 1M tokens. In addition, the thinking length of the new model has increased, recommending setting a longer thinking budget in highly complex reasoning tasks to fully unleash its potential.