Qwen3 Technical Report

The Qwen3 technical report presents the latest generation of the Qwen large language model family, comprising both dense and Mixture-of-Expert (MoE) architectures with parameter scales ranging from 0.6 to 235 billion. Its central innovation is the integration of a 'thinking' mode for complex, multi-step reasoning and a 'non-thinking' mode for rapid, context-driven responses within a single unified framework, removing the need to switch between separate chat-optimized and dedicated reasoning models. A thinking-budget mechanism lets users adaptively allocate inference compute to balance latency against performance according to task complexity.

By transferring knowledge from the flagship models, the authors substantially reduce the compute needed to build smaller yet competitive models. Empirically, Qwen3 achieves state-of-the-art results across benchmarks in code generation, mathematical reasoning, and agent tasks, competitive with larger MoE and proprietary models, while expanding multilingual coverage from 29 to 119 languages and dialects. All Qwen3 models are released publicly under the Apache 2.0 license to support reproducibility and community research.