# Qwen3 Technical Report

**Authors:** An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang  
**arXiv:** [2505.09388](https://arxiv.org/abs/2505.09388)  
**Date:** 2025-05-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.09388v1/

## Key points

- Qwen3 is a significant advancement in large language models (LLMs) with a focus on performance, efficiency, and multilingual capabilities.
- Qwen3 includes models with dense and Mixture-of-Expert (MoE) architectures, with parameter scales ranging from 0.6 to an impressive 235 billion.
- Qwen3 integrates thinking mode for complex reasoning tasks and non-thinking mode for rapid context-driven responses within a unified framework, enabling dynamic mode switching based on user queries or chat templates.
- Qwen3 introduces a thinking budget mechanism that allows adaptive allocation of computational resources during inference to balance latency and performance based on task complexity.
- Qwen3 achieves state-of-the-art results across diverse benchmarks including code generation, mathematical reasoning, agent tasks, competes favorably against larger MoE models and proprietary models in various domains.
- One key enhancement in Qwen3 is the expansion of multilingual support from 29 languages to 119 languages, enhancing global accessibility through improved cross-lingual understanding and generation capabilities.
- All models within the Qwen3 series are publicly accessible under Apache 2.0 licensing terms to promote reproducibility and foster community-driven research efforts.

## Summary

, , , , 

In the Qwen3 Technical Report, a team led by An Yang and including researchers such as Anfeng Li, Baosong Yang, Beichen Zhang, and many others introduces the latest iteration of the Qwen model family. Qwen3 is a significant advancement in large language models (LLMs), with a focus on enhancing performance, efficiency, and multilingual capabilities. The series includes models with both dense and Mixture-of-Expert (MoE) architectures, featuring parameter scales ranging from 0.6 to an impressive 235 billion. One notable innovation in Qwen3 is the integration of thinking mode for complex, multi-step reasoning tasks and non-thinking mode for rapid context-driven responses within a unified framework. This eliminates the need to switch between different models optimized for specific tasks like chat interactions or dedicated reasoning processes. Instead, dynamic mode switching based on user queries or chat templates is enabled. Additionally, Qwen3 introduces a thinking budget mechanism that allows users to allocate computational resources adaptively during inference to balance latency and performance based on task complexity. By leveraging knowledge from flagship models, the team behind Qwen3 has significantly reduced the computational resources required to build smaller-scale models while maintaining highly competitive performance levels. Empirical evaluations demonstrate that Qwen3 achieves state-of-the-art results across diverse benchmarks including code generation, mathematical reasoning, agent tasks, among others. It competes favorably against larger MoE models and proprietary models in various domains. Compared to its predecessor Qwen2.5, one of the key enhancements in Qwen3 is the expansion of multilingual support from 29 languages and dialects to an impressive 119 languages. This expansion enhances global accessibility through improved cross-lingual understanding and generation capabilities. To promote reproducibility and foster community-driven research and development efforts, all models within the Qwen3 series are made publicly accessible under Apache 2.0 licensing terms. The collaborative effort showcased in this technical report underscores a commitment to advancing language modeling capabilities while ensuring broad accessibility for researchers worldwide.