Qwen3 Technical Report
Report describing Qwen3's unified thinking and non-thinking modes, thinking-budget control, and 0.6B-235B dense and MoE models.
Explains the single-model design that removes switching between chat and reasoning models, the budget mechanism for inference compute and distillation of small models from large ones. The July 2507 releases later split the modes into separate checkpoints.
- Date
- Wednesday, 14 May 2025
- Lab
- Alibaba (Qwen)
- Kind
- paper
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Parameter range | 0.6B to 235B Dense and MoE; 60 authors | company |
arXiv 2505.09388 submitted 2025-05-14, two weeks after the models. The retreat from hybrid mode in July 2025 is visible only in the model cards; Qwen's stated reason was not found.
Sources
This record was checked against its sources on 6 October 2026. How we check