Alibaba releases Qwen3 open-weight models with hybrid thinking modes

Alibaba Qwen/Qwen models

Alibaba released Qwen3 as open-weight models under Apache 2.0: two mixture-of-experts models (235B total/22B active, 30B/3B) and six dense models from 0.6B to 32B. Each model switches between step-by-step reasoning and quick answers with adjustable computation budgets, trained on 36 trillion tokens across 119 languages.

• Eight models: two MoE (235B, 30B) and six dense (0.6B to 32B); all open-weight Apache 2.0
• Hybrid thinking modes: step-by-step reasoning or rapid responses with adjustable budgets
• Trained on 36 trillion tokens, 119 languages and dialects; context up to 128K tokens for larger models
• Small MoE model outcompetes Qwen-32B with 10x fewer activated parameters