On July 30, the Tongyi Qianwen team of Alibaba officially launched the latest member of the Qwen3 series model - Qwen3-30B-A3B-Instruct-2507. This new open-source model uses a non-thinking mode and only requires activating 3B parameters to match renowned closed-source models such as Gemini 2.5-Flash and GPT-4o in multiple core capabilities. In addition, Qwen3-30B-A3B-Instruct-2507 has achieved significant breakthroughs in multilingual support, user preference alignment, and long text processing.
The model is now fully open-sourced on the ModelScope community and HuggingFace platform. As a new version of the Qwen3-30B-A3B series, Qwen3-30B-A3B-Instruct-2507 not only inherits the advantages of the Tongyi Qianwen team in architectural innovation but also significantly reduces computational costs, striving to match the performance of trillion-parameter closed-source models.
According to official data, Qwen3-30B-A3B-Instruct-2507 performs well in multiple key benchmark tests, including mathematical reasoning (AIME25 test score of 61.3), code generation (LiveCodeBenchv6 score of 43.2), graduate-level physics and astronomy questions (GPQA test score of 70.4), and human preference alignment (Arena-Hard v2 score of 69), even surpassing GPT-4o in some metrics. This achievement marks a significant improvement in the model's comprehensive capabilities in logic reasoning, mathematics, science, and programming.
More notably, Qwen3-30B-A3B-Instruct-2507 also performs well in covering long-tail knowledge across multiple languages and better aligns with user preferences in subjective and open-ended tasks, generating higher-quality text to provide users with more valuable answers. At the same time, its long-text understanding capability has been improved to 256K, meaning users can handle more complex text content.
Since its launch in 2023, the Tongyi Qianwen series models have open-sourced over 200 models, with global downloads exceeding 300 million times, and more than 100,000 derivative models, demonstrating strong market influence and technological potential.