Researchers with less than $ 50 have made a free -reasoning artificial intelligence model

Over the past weeks, Chinese artificial intelligence Deepseek, built much lower than American models, has led to the fall of technology companies and markets around the world. Researchers at Stanford and Washington universities now claim in a new article that with less than $ 50 Succeed in building a model The reasonable artificial intelligence Free like O1 from Openai.

According to reports, the model that S1 It is called in tests that measure the ability of mathematics and coding, similar to advanced reasoning models such as O1 and Deepseek R1. The S1 is currently available along with the data and the code used to teach it in Github.

Build a Free S1 Artificial Intelligence Model

The researchers say in their article that they first developed a basic model and then adjusted it through a process called “distillation” used to extract the “reasoning” capabilities from another artificial intelligence model. According to them, Google’s Flash Thinking Experimental 2.0 Flash Thinking Experimental model is said.

According to the researchers, the S1 training with 16 Nvidia H100 graphics processors has lasted less than 30 minutes, and this model has achieved great performance in some artificial intelligence benchmarks.

The S1 -made research team has sought the easiest approach to achieving powerful performance in reasoning and “testing time”, the second allows the artificial intelligence model to think more before delivering the answer. Openai, of course, also made such improvements in its O1 model, and then Deepseek and other artificial intelligence laboratories have tried to use them through various techniques.

The article S1 shows that reasoning models can be distilled through a process called “Surveled Returned Settlement” (SFT) with a relatively small dataset. In this process, the artificial intelligence model is instructed to simulate specific behaviors in a dataset. The SFT process is said to be cheaper than the reinforcement learning method Deepseek has done to teach the R1 based on O1.