Qwen3
Qwen3 introduces hybrid reasoning combining thinking and non-thinking modes. The series includes six dense models and two MoE variants, all open-sourced globally and trained on 36 trillion tokens.
Open-source hybrid-reasoning AI from Alibaba
Qwen3 introduces hybrid reasoning combining thinking and non-thinking modes. The series includes six dense models and two MoE variants, all open-sourced globally and trained on 36 trillion tokens.
Qwen3 introduces hybrid reasoning combining thinking and non-thinking modes. The series includes six dense models and two MoE variants, all open-sourced globally and trained on 36 trillion tokens.
Pick Qwen3 when you need an open-source multilingual AI you can self-host or fine-tune — great for research and on-prem deployments.
Qwen3 supports 119 languages globally, making it one of the most linguistically broad open-source model families available.
Qwen3 supports both thinking mode (extended chain-of-thought for hard problems) and non-thinking mode (fast direct responses), which you can toggle per query.
Qwen3 weights are freely available on Hugging Face in multiple sizes from 0.6B to 235B parameters. All variants support commercial use under Alibaba's open licence.
Qwen3 introduces hybrid reasoning combining thinking and non-thinking modes. The series includes six dense models and two MoE variants, all open-sourced globally and trained on 36 trillion tokens.