S1 AI model explained: What is it and all you need to know about this ‘$50’ language model

artificial intelligence

Researchers from Stanford University and the University of Washington have developed an AI reasoning model called s1, which was trained for under $50 using a process called distillation. The model is based on Qwen2.5, an open-source language model from Alibaba Cloud, and was refined using outputs from Google’s Gemini 2.0 Flash Thinking Experimental model.

S1 AI model development details

The research team trained S1 on a small dataset of 1,000 questions instead of a larger dataset of 59,000 questions, determining that a smaller dataset was sufficient for strong reasoning performance. The model was refined using distillation, a technique that extracts reasoning patterns from larger AI models.

To train S1, researchers used 16 Nvidia H100 GPUs, a significantly lower compute cost than many large-scale AI models. The training process included supervised fine-tuning (SFT), where the model was explicitly taught to mimic patterns in the dataset.

Key differences between s1 and ChatGPT

The S1 model differs from OpenAI’s ChatGPT in its development process, dataset size, and approach to reasoning. Unlike ChatGPT, which is trained on large-scale datasets requiring extensive computing resources, s1 was trained using a much smaller dataset and fewer computing resources.

Another difference is S1’s use of test-time scaling, which allows the model to spend more time refining responses. Researchers implemented a technique called budget forcing, which forces the model to continue its reasoning process by appending the word “Wait” multiple times, encouraging it to reassess its answers.

The S1 model has been compared to OpenAI’s o1 and DeepSeek’s R1 reasoning models. According to the researchers, s1 outperforms o1-preview by up to 27% on mathematical reasoning tasks. DeepSeek’s R1 model was also developed using distillation techniques, similar to S1.

The rise of models like s1 suggests that advanced AI reasoning capabilities can be achieved with smaller datasets and lower training costs. This approach challenges the current model development strategies of companies like OpenAI, Microsoft, Google, and Meta, which rely on high-budget AI training and large-scale data centres.

Availability of S1

The researchers have made the s1 model, along with its training data and code, available on GitHub. The project aims to contribute to further research in cost-effective AI reasoning models.