Alibaba Cloud has confirmed that research teams from Stanford University and the University of Washington, led by Fei-Fei Li, have developed an artificial intelligence reasoning model named s1 based on the Alibaba Tongyi Qwen (Qwen) model. This model performs comparably to cutting-edge reasoning models such as OpenAI's o1 and DeepSeek's R1 in mathematics and coding ability tests, even surpassing o1-preview by 27% on competitive math problems. The training process utilized 16 H100 GPUs, took 26 minutes, and incurred cloud computing costs of less than $50.

PANews reported on February 6 that according to Sina Technology, Fei-Fei Li and other researchers from Stanford University and the University of Washington trained an artificial intelligence reasoning model called s1 with a cloud computing fee of less than $50. The model's performance in math and coding ability tests was similar to that of cutting-edge reasoning models such as OpenAI's o1 and Depsek's R1, which attracted widespread attention. However, it was soon reported that the s1 model was "not trained from scratch" and that its base model was the "Alitong Yi Qianwen (Qwen) model." In response, the reporter asked Alibaba Cloud for confirmation, and Alibaba Cloud confirmed the news and responded: "They used the Alitong Yi Qianwen Qwen2.5-32B-Instruct open source model as the base, supervised fine-tuning on 16 H100 GPUs for 26 minutes, and trained a new model s1-32B, which achieved results comparable to the math and coding capabilities of cutting-edge reasoning models such as OpenAI's o1 and DeepSeek's R1, and even performed 27% higher than o1-preview on competition math problems."

Alibaba Cloud confirms that the S1 model developed by Fei-Fei Li's team is trained based on Qwen.