The Qwen3 Technical Report introduces a new series of open-weight large language models ranging from 0.6 to 235 billion parameters, featuring both dense and Mixture-of-Experts architectures. A key innovation is the integration of "thinking mode" for complex reasoning and "non-thinking mode" for rapid responses into a single framework, allowing dynamic switching without model alternation.
- The flagship Qwen3-235B-A22B model utilizes 22 billion activated parameters per token to balance performance and inference efficiency.
- Multilingual support expands from 29 to 119 languages and dialects, trained on 36 trillion tokens including synthetic data generated by previous Qwen models.
- The series implements a "thinking budget" mechanism for adaptive computational resource allocation during inference.
- Post-training employs multi-stage reinforcement learning and strong-to-weak distillation to align the models with human preferences.
The authors consider these advancements important because they enable competitive performance against proprietary models while providing users with fine-grained control over reasoning effort and latency.