Alibaba releases the Qwen2.5 family of open-weight models

Alibaba Qwen/Qwen models

Alibaba released Qwen2.5, open-weight language models from 0.5B to 72B parameters trained on up to 18 trillion tokens, alongside specialised Qwen2.5-Coder and Qwen2.5-Math models. Most released under Apache 2.0; the 3B and 72B variants under separate licences. Coder trained on 5.5 trillion tokens of code; Math supports reasoning in Chinese and English.

• Qwen2.5: 0.5B to 72B parameters; pretrained on 18 trillion tokens; context up to 128K, generate 8K tokens
• Qwen2.5-Coder: 1.5B, 7B, 32B; trained on 5.5 trillion code tokens; HumanEval 85+
• Qwen2.5-Math: 1.5B, 7B, 72B; multilingual (Chinese, English); supports multiple reasoning approaches
• Most under Apache 2.0; 3B and 72B variants under separate terms; 29+ language support