Ali Tongyi Thousand Questions Launches New Reasoning Model Qwen3-30B-A3B-Thinking-2507 with Significant Improvements in Several Capabilities

July 31, 2012 - AliThousand Questions on TongyiToday announced the launch of the newinference model Qwen3-30B-A3B-Thinking-2507. Compared to the Qwen3-30-A3B model, which was open sourced on April 29, the new model offers significant improvements in reasoning, generalization, and context length:

• The new model scored 85.0 on AIME25, which focuses on math proficiency, and 66.0 on LiveCodeBench v6, a test of code proficiency.Both core reasoning abilities exceed Gemini 2.5-Flash (thinking), Qwen3-235B-A22B (thinking); the knowledge level of the new model (GPQA, MMLU-Pro) has also improved significantly from the previous version.
• In the generalized competency measures of WritingBench, Agent Competency (BFCL-v3), Multi-Round Dialogue, and Multi-Language Instruction Following (MultiIF), Qwen3-30B-A3B-Thinking-2507 outperforms Gemini2.5-Flash (thinking), Qwen3-235B-A22B (thinking), Qwen3-235B-A22B (thinking), and Qwen3-235B-A22B (thinking). 235B-A22B (thinking).
• longer contextual understanding, with native support for 256K tokens.Scalable to 1M tokens.

also,The thinking length of the new model has also been increased by, it is officially recommended to set a longer thinking budget in highly complex reasoning tasks to realize its full potential.

Officially, Qwen3-30B-A3B-Thinking-2507 has been open-sourced in the Magic Hitch community and HuggingFace, and its lightweight size makes it easy to deploy consumer-grade hardware locally; at the same time, it has also launched a new model on Qwen Chat.