A team of AI researchers from Stanford and the University of Washington has successfully trained an AI reasoning model for less than $50 using cloud computing credits. Named s1, the model performs at a level comparable to advanced AI reasoning models, including OpenAI’s o1 and DeepSeek’s R1, in math and coding tasks. The research team has made s1 openly available on GitHub, along with the data and training code.

To develop s1, the researchers used an existing small AI model from Alibaba-owned Qwen and fine-tuned it through a process called distillation. This method extracts reasoning abilities from a larger AI model by training on its answers. The team distilled s1 from Google’s Gemini 2.0 Flash Thinking Experimental, which is freely accessible through Google AI Studio, though with daily usage limits. A similar distillation approach was used last month by Berkeley researchers to create an AI reasoning model for around $450.

Training s1 required just 1,000 carefully selected questions, each paired with an answer and a detailed reasoning process from Gemini 2.0 Flash Thinking Experimental. The entire training process was completed in under 30 minutes using 16 Nvidia H100 GPUs. Niklas Muennighoff, a Stanford researcher involved in the project, stated that renting the necessary computing power today would cost about $20.

The success of s1 raises concerns about AI model commoditization. If a high-performing model can be replicated at minimal cost, it challenges the competitive advantage of major AI firms. OpenAI has already accused DeepSeek of improperly collecting data from its API to train its own reasoning model, R1. This highlights growing concerns over how proprietary AI models can be replicated through distillation techniques.

One notable finding from the research is that adding the word "wait" during s1’s reasoning process improved accuracy. By instructing the model to pause before responding, researchers found that it could generate more precise answers.

Despite the affordability of distillation, major tech companies are still investing heavily in AI development. In 2025, Meta, Google, and Microsoft plan to spend hundreds of billions of dollars on AI infrastructure to train next-generation models. While distillation offers a cost-effective way to replicate existing AI capabilities, it does not create entirely new models that outperform current industry leaders.

SOURCE: TECHCRUNCH | PHOTO: SHUTTERSTOCK

This article was created with AI assistance.