Open-Source Approaches to Artificial Intelligence: A New Frontier
Open-source approaches continue to show promise in democratizing artificial intelligence (AI).
NovaSky’s Sky-T1-32B-Preview
On Friday, the NovaSky research team at UC Berkeley released a new reasoning model, Sky-T1-32B-Preview, that performs comparably to OpenAI’s o1-preview — only it’s open source and was built in just 19 hours for under $450 using eight Nvidia H100 GPUs.
The Best of Both Worlds: Open-Source AI Models
The team developed Sky-T1 by fine-tuning Alibaba’s Qwen2.5-32-Instruct and trained it on data generated with QwQ-32B-Preview, another open-source model comparable to o1-preview. Using synthetic training data can help lower costs.
Data Preparation: The Key to Success
"We curate the data mixture to cover diverse domains that require reasoning, and a reject sampling procedure to improve the data quality. We then rewrite QwQ traces with GPT-4o-mini into a well-formatted version, inspired by Still-2, to improve data quality and ease parsing," the team says of their data preparation process in the blog.
Outperforming OpenAI’s o1-Preview
The model performed at or above o1-preview’s level on math and coding benchmarks but did not surpass o1 on the graduate-level benchmark GPQA-Diamond, which includes more advanced physics-related questions. NovaSky open-sourced all parts of the model, including weights, data, infrastructure, and technical details.
A More Affordable Reasoning Model
The relatively short 19-hour training time means Sky-T1 cost just $450 to build, according to Lambda Cloud pricing, the team clarifies in the blog post. Considering GPT-4 used a suspected $78 million in compute, it is no small feat to present an example of a more affordable reasoning model that can be replicated by academic and open-source groups that lack OpenAI’s funding.
Conclusion
The continued development of open-source AI models like Sky-T1-32B-Preview holds great promise for democratizing AI and creating a more even playing field for smaller labs, nonprofits, and other entities to develop competitive models. As the field of AI continues to evolve, it will be exciting to see how these open-source models can be used to drive innovation and progress.
FAQs
Q: What is the significance of Sky-T1-32B-Preview?
A: Sky-T1-32B-Preview is an open-source reasoning model that performs comparably to OpenAI’s o1-preview, built in just 19 hours for under $450 using eight Nvidia H100 GPUs.
Q: How was Sky-T1-32B-Preview developed?
A: The team developed Sky-T1 by fine-tuning Alibaba’s Qwen2.5-32-Instruct and trained it on data generated with QwQ-32B-Preview, another open-source model comparable to o1-preview.
Q: Is Sky-T1-32B-Preview available for use?
A: Yes, NovaSky open-sourced all parts of the model, including weights, data, infrastructure, and technical details.
Q: What are the potential applications of Sky-T1-32B-Preview?
A: The potential applications of Sky-T1-32B-Preview are vast, including but not limited to, natural language processing, computer vision, and robotics.
