A team of researchers at Stanford University and the University of Washington have developed an open-source AI reasoning model that purportedly performs on par with OpenAI o1 and DeepSeek R1 models. This model, known as the s1, was revealed to be built for less than US$50 (RM221.65) of cloud compute credits and required only 26 minutes to train.
In a research paper published last Friday, the team explained that the s1 was developed using an “off-the-shelf” model from Alibaba-owned Chinese AI lab Qwen. The model was trained on 1,000 questions on 16 NVIDIA H100 GPUs using a process called “supervised fine-tuning”.
The team also controlled the computing time for s1 by forcing it to generate an answer when it spent too much time on a problem. Conversely, the researchers could also tell the model to wait and extend its thinking time, which produced more accurate results.
This is not the first low-cost open-source reasoning model. In January, researchers from UC Berkeley released Sky-T1, a model which cost $450. Other reasoning models released within this same price range include rStar-Math from Microsoft Asia researchers and Ai2’s Tulu 3, according to Mashable.
The rise of cheap AI models is no surprise as the industry seeks to save on computing costs, with more low-cost projects likely to appear in the near future. This sudden surge is credited to Chinese startup DeepSeek, which took the world by storm when it introduced its own AI model last month. According to the company, the cost of development was below US$6 million (RM26.59 million), which is far less when compared to the billions spent by western AI firms.
(Source: Mashable)
Nurul Kamil contributed to this article.
Follow us on Instagram, Facebook, Twitter or Telegram for more updates and breaking news.
