Alibaba Unveils QwQ-32B, a New Open-Source AI Model Rivaling DeepSeek-R1

Alibaba Group's shares surged after releasing QwQ-32B, a new open-source AI reasoning model. The model rivals DeepSeek-R1 in performance but is significantly smaller, with 32.5 billion parameters compared to R1's 671 billion (though R1 only activates 37 billion at a time). QwQ-32B, based on the Transformer architecture, incorporates rotary positional encoding for improved text understanding. It can process prompts with up to 131,072 tokens and excels in coding, math, and external application tasks. Developed using reinforcement learning, QwQ-32B outperformed R1 in benchmarks measuring external system interaction, question-answering, and output alignment. This release follows Alibaba's commitment to invest $53 billion in AI infrastructure over the next three years. Other Chinese tech giants like Tencent are also prioritizing LLM development, highlighting the growing AI race in China.