Tongyi Qwen3’s new non-thinking model has been updated, and its multiple capabilities have reached the level of Gemini2.5 and GPT-4o

On July 30, Alibaba Tongyi Qianwen team announced the launch of the latest version of the Qwen3 series model - Qwen3-30B-A3B-Instruct-2507. This non-thinking mode open source model only needs to activate 3B parameters to achieve performance levels comparable to closed source models such as Gemini 2.5-Flash (non-thinking) and GPT-4o in multiple core capabilities. It also achieves breakthrough improvements in areas such as multi-language coverage, user preference alignment, and long text processing.

Currently, the model has been fully open sourced in the ModelScope community and the HuggingFace platform.

As a new version of the Qwen3-30B-A3B model, Qwen3-30B-A3B-Instruct-2507 continues the advantages of the Tongyi Qianwen team in architectural innovation. While significantly reducing computing costs, it achieves performance benchmarking with hundreds of billions of closed-source models. Official data shows that the model performs outstandingly in key benchmark tests such as mathematical reasoning (AIME25 test score 61.3), code generation (LiveCodeBenchv6 score 43.2), graduate-level physics and astronomy problems (GPQA test score 70.4), human preference alignment (Arena-Hard v2 score 69), function calling ability (BFCL-v3 score 65.1), and some indicators even surpass GPT-4o.

At the general ability level, the model comprehensively strengthens instruction following, logical reasoning, mathematics, science, programming and tool use. In terms of long-tail knowledge coverage in multiple languages, the model has made significant progress; in subjective and open tasks, the new model is further closely aligned with user preferences, can generate higher-quality text, and provide users with more helpful answers; the long text understanding ability has been improved to 256K.

Tongyi Qianwen is a large language model series independently developed by Alibaba Damo Academy. Since its release in 2023, more than 200 models have been open sourced, with global downloads exceeding 300 million times, and the number of derivative models exceeding 100,000.