How To Train A ‘Reasoning’ AI Model For Less Than $450?
The AI world has just been shaken, stirred, and served with a twist of affordability. The NovaSky team from UC Berkeley’s Sky Computing Lab has dropped a bombshell on the AI community with the release of Sky-T1-32B-Preview, an open-source reasoning AI model that can be trained for the jaw-dropping cost of—wait for it—less than $450. Yes, you read that right. In a world where training AI models often costs millions of dollars, this is like finding a Tesla for the price of a used bicycle.
Let’s dive into the details of this groundbreaking development, explore how it compares to other AI models, and understand why this could be a game-changer for the industry.
The Birth of Sky-T1: A Budget-Friendly Marvel
The Sky-T1-32B-Preview is the brainchild of the NovaSky research team at UC Berkeley’s Sky Computing Lab. Released on January 11, 2025, this model is being hailed as the first truly open-source inference AI model. Not only is the model itself open-source, but the team has also generously shared the dataset and training code used to develop it. This means anyone with the right hardware can replicate the model from scratch.
The NovaSky team achieved this feat by leveraging synthetic training data. They used Alibaba’s QwQ-32B-Preview model to generate the initial training data, which was then curated and refined using OpenAI’s GPT-4o-mini. The final training process took just 19 hours on a rack of 8 Nvidia H100 GPUs, a setup that many AI enthusiasts might already have access to.
