Today, Alibaba Tongyi Qianwen announced the release of a new smaller model——Qwen3-4B-Instruct-2507 and Qwen3-4B-Thinking-2507.At present, the new model has been officially open sourced in the Magic Community and HuggingFace. According to reports,In the non-inference field, Qwen3-4B-Instruct-2507 comprehensively surpasses the closed-source GPT4.1-Nano.
In the field of reasoning, the Qwen3-4B-Thinking-2507 is even comparable to the mid-sized Qwen3-30B-A3B (thinking).
Officials stated that the 2507 version of the Qwen3-4B model is particularly friendly to the deployment of end-side hardware such as mobile phones.
The following are the core highlights of the model
Qwen3-4B-Instruct-2507
The general capabilities have been greatly improved, surpassing the commercial closed-source small-size model GPT-4.1-nano, and approaching the performance of the medium-sized Qwen3-30B-A3B (non-thinking).
The new model covers long-tail knowledge in more languages, enhances human preference alignment in subjective and open-ended tasks, and can provide responses that are more in line with people's needs.
Contextual understanding extends to 256K, and small models can also handle long texts.
Qwen3-4B-Thinking-2507
The reasoning ability has been greatly enhanced, with AIME25 scoring as high as 81.3. The reasoning performance of Qwen3-4B-Thinking-2507 is comparable to the mid-range model Qwen3-30B-Thinking.
Especially in the AIME25 assessment, which focuses on mathematical ability, it scored 81.3 points with 4B parameters.
Agent scores are off the charts, and all relevant reviews surpass the larger Qwen3-30B-Thinking model.
The contextual understanding ability of 256K tokens supports more complex document analysis, long-form content generation, cross-paragraph reasoning and other scenarios.
