Open Source AI Model Rivals OpenAI's Reasoning Model for Under $50

Affordable AI Reasoning Model Developed by Researchers

A new research paper, published last Friday, details how AI researchers from Stanford University and the University of Washington successfully trained an AI model capable of “reasoning” for under $50 in cloud computing credits.

This model, designated s1, demonstrates performance comparable to state-of-the-art reasoning models like OpenAI’s o1 and DeepSeek’s R1 when evaluated on tests assessing mathematical and coding skills. The s1 model, along with the associated data and training code, is publicly accessible on GitHub.

Model Training and Distillation

The team developing s1 began with a readily available base model. They then employed a technique called distillation to imbue it with “reasoning” abilities. Distillation involves training the model on the outputs of another, more complex AI model, effectively extracting its reasoning processes.

Specifically, s1 was distilled from Google’s Gemini 2.0 Flash Thinking Experimental reasoning model. This approach mirrors that used by researchers at Berkeley last month, who created a similar AI reasoning model for approximately $450.

Implications for AI Commoditization

The ability for a small research team, without substantial financial backing, to achieve such innovation in the AI field is noteworthy. However, s1 also raises important questions regarding the potential commoditization of AI models.

The ease with which a multi-million-dollar model can be replicated with minimal expenditure prompts consideration of competitive advantages within the AI landscape.

Large AI laboratories have expressed concern. OpenAI, for instance, has accused DeepSeek of utilizing data improperly obtained from its API for model distillation purposes.

Focus on Reasoning Performance and Scaling

The s1 researchers aimed to identify the most straightforward method for achieving robust reasoning performance and “test-time scaling” – enabling an AI model to dedicate more processing time to a question before providing an answer. These were key advancements in OpenAI’s o1 model, which DeepSeek and other labs have attempted to replicate.

The research suggests that reasoning models can be effectively distilled using a relatively small dataset through supervised fine-tuning (SFT). SFT involves explicitly instructing the AI model to emulate specific behaviors present within a dataset.

SFT generally proves less expensive than the large-scale reinforcement learning approach utilized by DeepSeek in the training of its R1 model, a competitor to OpenAI’s o1.

Access to Google’s Gemini and Model Foundation

Google provides free access to Gemini 2.0 Flash Thinking Experimental via its Google AI Studio platform, although daily usage is subject to rate limits. However, Google’s terms of service prohibit reverse-engineering its models to create competing AI services. A request for comment has been sent to Google.

S1 is built upon a smaller, pre-existing AI model developed by Qwen, a Chinese AI lab owned by Alibaba, and is freely available for download.

Training and Performance Details

To train s1, the researchers curated a dataset of 1,000 carefully selected questions, paired with corresponding answers and the reasoning process behind each answer, derived from Google’s Gemini 2.0 Flash Thinking Experimental.

Training s1 took less than 30 minutes utilizing 16 Nvidia H100 GPUs. According to the researchers, the necessary compute resources could currently be rented for around $20, as stated by Niklas Muennighoff, a Stanford researcher involved in the project.

Enhancing Reasoning with a Simple Prompt

The researchers discovered a simple technique to encourage s1 to verify its work and extend its reasoning time: the inclusion of the word “wait” during the reasoning process. This prompt led to slightly more accurate responses, as detailed in the research paper.

Future Investment in AI Infrastructure

In 2025, Meta, Google, and Microsoft are planning substantial investments, totaling hundreds of billions of dollars, in AI infrastructure, partially dedicated to the development of next-generation AI models.

Such significant investment may remain crucial for driving further advancements in AI innovation. While distillation has proven effective for cost-effectively replicating existing AI capabilities, it does not inherently create entirely new models that surpass current performance levels.