A new reasoning AI model, Sky-T1, has been released by the NovaSky research team at UC Berkeley’s Sky Computing Lab.
The team’s breakthrough demonstrates that it is now possible to create high-level reasoning models affordably, with Sky-T1 trained for less than $450.
Affordable AI Training
Sky-T1-32B-Preview is the first open-source reasoning AI model that can be replicated from scratch. The team has made the training data and code available for others to use, showing how reasoning models can be developed on a budget.
Previously, training a model with similar performance could cost millions of dollars. However, with synthetic training data — data generated by other AI models — the cost of training has dropped significantly.
For comparison, Palmyra X 004, another AI model trained mostly on synthetic data, cost $700,000 to develop. Sky-T1’s low cost is a major step forward for the AI community.
What Makes Sky-T1 Special?
Sky-T1 is a reasoning model, which means it can fact-check its own answers. This self-checking ability makes it more reliable in fields like physics, science, and mathematics.
While reasoning models take longer to find solutions (usually seconds to minutes), they tend to be more accurate than regular AI models.
Sky-T1-32B-Preview was trained using 8 Nvidia H100 GPUs over 19 hours. The model outperforms an earlier version of OpenAI’s o1 on math and coding challenges, such as the MATH500 and LiveCodeBench tests. However, it still lags behind o1 on more complex topics like physics, biology, and chemistry.
The Future of Sky-T1 and Reasoning AI
The NovaSky team is not stopping with Sky-T1. They plan to develop even more efficient models that retain strong reasoning capabilities while improving accuracy and performance.
The team is excited to continue their work on advanced reasoning models, making them more accessible and affordable for everyone.
