A team of AI researchers from Stanford and the University of Washington has developed an AI reasoning model, s1, for under $50 in cloud computing costs. The model, which performs comparably to advanced reasoning models like OpenAI’s o1 and DeepSeek’s R1 in math and coding tests, is available on GitHub along with its training data and code.
The researchers used a process called distillation, where an AI model is trained on another model’s responses to extract its reasoning abilities. In this case, s1 was distilled from Google’s Gemini 2.0 Flash Thinking Experimental.
The project highlights how AI capabilities can be replicated at a fraction of the cost of traditional large-scale models. This raises concerns for major AI companies investing billions in AI development. OpenAI, for instance, has accused DeepSeek of using its API data for similar distillation purposes.
The researchers behind s1 aimed to explore the simplest way to achieve strong reasoning performance and test-time scaling, a method that allows AI models to think more before delivering an answer.
S1 was trained using a small dataset of 1,000 carefully curated questions with detailed answers and reasoning steps from Google’s model. The training process took less than 30 minutes on 16 Nvidia H100 GPUs, with one researcher estimating the required compute cost at around $20.
A notable technique they used involved inserting the word “wait” into prompts, encouraging the model to take extra time for reasoning, leading to slightly more accurate answers.Despite the success of distillation, the researchers acknowledge that while it allows for cost-effective replication of AI capabilities, it does not drive groundbreaking advancements in AI. Meanwhile, major tech firms, including Meta, Google, and Microsoft, continue to invest heavily in AI infrastructure, aiming to push the boundaries of innovation with large-scale models.
