Researchers from Stanford University and the University of Washington in the United States revealed in a recently published research paper that they trained a cutting-edge artificial intelligence (AI) inference model at an expense of less than $50.
The researchers explained that the model named "s1" showed performance comparable to OpenAI's "o1" and DeepSeek's "R1" in mathematics and coding ability tests. o1 is the inference model that OpenAI first released last year. R1 is the inference model that DeepSeek introduced last month, and it is known to have similar performance to o1.
The data and code used for training the s1 model have been made public on the web-based platform GitHub, which helps developers store and manage their code and files.
The researchers explained that during the development of s1, they performed fine-tuning through a technical process called "distillation." Distillation refers to the AI model using the output results of another model for training purposes to develop similar functionalities.
s1 was distilled from Google's latest AI model, Gemini 2.0 Flash Thinking Experimental. The researchers noted that "training s1 took less than 30 minutes using NVIDIA's advanced AI chip, the H100 graphics processing unit (GPU)," and added, "the total expense was less than $50." They also mentioned that the computing performance needed to train s1 could be utilized for about $20.