Researchers created an open rival to OpenAI’s o1 ‘reasoning’ model for under $50

AI researchers at Stanford and Washington University were able to train an AI “logic” model below $ 50 on cloud compots credits, according to a new Research paper Released last Friday.

The model, known as the S1, performs the same with the Cutting-Edge logic models like OpenAI and 1 and DEPSEC R1 in the tests of mathematics and coding capabilities. S1 is the model Available at GithubThis is including data and code used for training.

The team behind the S1 says that they started with an off-the-shelf base model, then it was finely tune in with the plain, training its answers to another AI model to lift the “logic” capacity. Researchers say that S1 is one of the reasonable models of Google, Gemi 2.0 flash thoughts have been emitted from experimental. Patna used in the same approach to the researchers in Berkeley Create an AI argument model for about 450 dollars last monthThe

To some, it is exciting that some researchers can still be invented in the AI ​​space without a few million dollars behind them. However, the S1 raises real questions about the producting of AI models. If someone can closely replicate the multi -million dollar model with a relative pocket change, where is the moat?

Surprisingly, big AI labs are not happy. Openi DIPSEC complained of harvesting data from its API ModelThe

Researchers behind S1 wanted to look for the easiest method for achieving “Test-time scaling” or allowed an AI model to think more before answering an question. They were some of the OPENCIES and 1’s Testaire, which DEPSEC and other AI labs have tried to replicate through various techniques.

The S1 paper suggests that the rational models are monitored with a process called Fine-Tuning (SFT) can be used with a relatively small datasate, where an AI model is clearly instructed to duplicate the specific behavior in a dataset. DEPSEC for the answer training of SFT Openai and 1, 1, the larger size is cheaper than the learning method of reinforcement.

Google’s Google AI Studio platform, with the daily rate limit, provides the 2.5 flash Thinking experimentally. However, its conditions prohibit its models opposite engineering to develop services that compete with Google’s own AI offers. We arrived on Google for comments.

S1 Alibaba-owned Chinese AI Lab is based on a small, off-shield AI model, which is available for free download. For S1 training, researchers only created a dataset of curates with 1000 caution, as well as answers to those questions, as well as the “Thought” process behind each answer from Google’s Gemi 2.0 flash thought experiment.

After the S1 training, which took less than 30 minutes using 16 Nvidia H 100 GPU, S1 achieved strong performance in some AI benchmarks, researchers said. Stanford researcher Niklas Muinighf, working on the project, told TechCrunch that he could rent the required count for about $ 20 today.

Researchers used a nifty technique to double-check the work of S1 and extend its “thoughts” time: they asked to wait. The word “wait” during the reason S1’s argument helped the model to come to a bit more accurate answer according to the paper.

In 2025, Meta, Google and Microsoft Plan a few billion dollars investment in AI infrastructureWhich will be partially training for the next generation of AI models. That level of investment to push the envelope of AI innovation may still be necessary. Distribution AI model’s ability has shown as a good method to recover cheap, but it does not create new AI models than what is available today.