ArchitectureTrainingReinforcement LearningEvaluation
Qwen3 Technical Report
Alibaba (Qwen Team)·May 14, 2025
Qwen Team
View on arXivTL;DR
A family of dense and MoE LLMs (0.6B–235B) that unifies a multi-step “thinking” mode and a fast “non-thinking” mode in one model with a tunable thinking budget, and expands multilingual support to 119 languages. Apache 2.0.
Why it matters
One of the most-used open-weight families of the period; it popularized the user-controllable “thinking budget” and the unified hybrid-reasoning interface that many later models copied.