A new AI model, Sky-T1-32B-Preview, developed by the NovaSky research team at UC Berkeley’s Sky Computing Lab, is making waves for its affordability and performance.
Released on Friday, Sky-T1 is the first open-source reasoning AI model that allows full replication, as the team made the training dataset and necessary code publicly available.
Remarkably, the model was trained for less than $450, demonstrating that high-level reasoning AI capabilities can be developed efficiently and cost-effectively.
Just a few years ago, training models with similar performance would have cost millions of dollars, but advances in synthetic training data and improved AI methods have made this possible.
Sky-T1-32B-Preview has shown strong results, surpassing an earlier version of OpenAI’s o1 in several key benchmarks. The model excelled in MATH500, a set of challenging math problems, and outperformed o1 in a coding challenge from LiveCodeBench.
However, Sky-T1 fell short when tested on GPQA-Diamond, which includes questions in physics, biology, and chemistry—subjects that require more advanced knowledge.
While Sky-T1 shows promise, OpenAI’s upcoming o3 release is expected to offer even better performance in reasoning tasks.
Despite this, Sky-T1 marks a significant milestone for affordable AI development, thanks to the strategic use of synthetic data and cost-efficient training.
The NovaSky team used Alibaba’s QwQ-32B-Preview model to generate the initial training data and then refined it using OpenAI’s GPT-4o-mini.
The model, with 32 billion parameters, was trained in just 19 hours using eight Nvidia H100 GPUs. Although reasoning models like Sky-T1 take a little longer to process answers compared to non-reasoning models, they are more reliable for tasks involving complex reasoning, such as scientific calculations.
This makes Sky-T1 a valuable tool for industries where precision matters, including physics, science, and mathematics. Looking ahead, the team is committed to continuing the development of open-source AI models with strong reasoning capabilities.