Researchers train AI model that rivals DeepSeek and OpenAI’s o1 for just $50

A team of researchers from Stanford University and the University of Washington has reportedly trained an AI reasoning model for just AED 184 (about $50).

The AI model, called s1, is said to be on par with some of the top industry models, like OpenAI’s o1 and China-based DeepSeek’s R1, especially when it comes to math and coding skills. They’ve now made s1, along with all the data and code used to train it, available on GitHub for anyone to test out.

Researchers say the S1 was developed through a process known as distillation, which is essentially like a teacher passing knowledge down to students. They started with an off-the-shelf AI model from Alibaba’s Qwen and then fine-tuned it using Google’s latest Gemini 2.0 Flash Experimental, a free tool, though with a daily usage limit.

The platform lets developers tap into Gemini’s full AI potential, including multimodal capabilities like text, audio, and visual processing, making it a powerful tool for training advanced models like S1.

Stanford presents:

s1: Simple test-time scaling

– Seeks the simplest approach to achieve test-time scaling and strong reasoning performance
– Exceeds o1-preview on competition math questions by up to 27% (MATH and AIME24)
– Model, data, and code are open-source pic.twitter.com/6rpGEid3ei

— Aran Komatsuzaki (@arankomatsuzaki) February 3, 2025

The researchers trained the S1 model using 1,000 carefully selected questions and answers, all based on Gemini 2.0’s responses, not just the answers themselves but also the reasoning behind them. To fine-tune the model, they used Supervised Fine-Tuning (SFT), a method that is faster and less time-consuming compared to DeepSeek’s Reinforcement Learning (RL), which basically forces the AI to figure things out on its own. What’s more, they managed to train S1 in just 30 minutes using 16 Nvidia GPUs.

The group states: “Recent advances in reasoning, such as OpenAI’s o1 and DeepSeek’s r1, lack transparency, limiting broader research progress. Our work aims to push the frontier of reasoning in a fully open manner, fostering innovation and collaboration to accelerate advancements that ultimately benefit society.”

That said, OpenAI accused DeepSeek of having trained its own AI open-source model using the company’s proprietary technology. While developers are allowed to use distillation, an OpenAI spokesperson told ReadWrite: “We are aware of and reviewing indications that DeepSeek may have inappropriately distilled our models, and will share information as we know more.

“We take aggressive, proactive countermeasures to protect our technology and will continue working closely with the U.S. government to protect the most capable models being built here.”

Last month, DeepSeek revealed that their AI model was trained for less than $20 million, sending Silicon Valley into a frenzy. The announcement triggered one of the biggest stock market crashes in US history, wiping nearly $1 trillion off the market. Nvidia, a company that had been soaring in value due to its GPUs being crucial for AI development, saw almost $600 billion erased from its worth.

Featured image: Canva

The post Researchers train AI model that rivals DeepSeek and OpenAI’s o1 for just $50 appeared first on ReadWrite.