Researchers from Stanford University and Washington University have created a new AI model called S1-32B that is similar to OpenAI's o1 model. They wanted to understand how the o1 model works. The dataset for the AI model was created using Gemini Flash Thinking. The AI model was not built from scratch but was based on another model called Qwen2.5-32B-Instruct. During the development of the AI model, the researchers used the Gemini Flash Thinking API to generate reasoning traces and responses. They created a dataset called s1K with high-quality questions and reasoning traces. The model was fine-tuned using basic parameters and took 26 minutes to train on 16 Nvidia H100 GPUs. The researchers found that they could manipulate the model's inference time by adding XML tags. They experimented with different phrases but found the best results with a "wait" tag. This tag allowed the model to think beyond the usual inference time. By using this method, the researchers were able to bring the model close to the performance of OpenAI's o1 model. The researchers were able to create the S1-32B AI model for under $50, showing that creating powerful AI models can be done at a low cost. They believe that their method may be similar to how OpenAI fine-tunes its reasoning models. The researchers hope that their work will help others understand AI models better and improve their performance. Overall, the researchers have created a low-cost open-source AI model that can analyze and understand complex reasoning tasks, similar to OpenAI's advanced models. This model has the potential to open up new possibilities for AI development and research.