Investigators created a ai model of AI in par nor O1’s O1 less than $ 50

Floods open to build thinking models AI is cheap.

Stanford and the University of Washington Developments Develop a model that acts similarly with the opening model O1 and DEPSEEK R1 Mathematics and codes – less than $ 50 of Cloud Credits.

In addition, the model was trained for only 1,000 questions, and only took 26 minutes and 16 Nvidia H100 GPUS. The Stanford’s Researcher Unikhas Muennightf is an e-mail to reduce costs that cost is based on GPU Runtime and H100 GPUS number used.

The AI ​​of the AI ​​of the Drunk is about the new methods of previous training process that can execute the computers, as proven by the DeepseeEek’s impact. In addition, developers have been able to build over AI models at existing or missing expenses, using the APIs, open source access, and a closed source models by installing their more money.

According to a published group research on the previous Friday, S1 was trained in the dataset that contained “carefully selected questions in accordance with the consultants and the answers filed by Gemini test.” The Gemini Thembema’s test model is accessible for daily limits through AI Studio. While it is a closed source model, that clearly not stopped researchers to use their answers.

Next, researchers have used the “shelf” with a chosen attitude from these fatherly, qwen, and we make the best direction of its selected data. The team later created a budget of the token to control the amount of model assessment. If S1 exceeds the budget from thought tokens, it was decided and forced to produce anything that came from. If researchers want a model to use the “Time-Time Complete”, they simply told the model that they were “waiting,” which increased its time thinking and led to proper results.

By managing the amount of time and computing computes used for a problem, researchers are able to show how the group increases lead to better performance.

S1 is one example of open source models developed for flagship fractions from Google and Openai. In January, UC Berkeley researchers have issued a vacant model called $ 450, “indicating that it is possible to multiply high-score skills,” by their mail. There is a model of RSTAR-MATE-Math’s open consultation

As high-quality models are available and cheap, we begin to see a change of energy from a few Ai Heats, for many.