Researchers at the University of California, Berkeley, have unveiled an open-source AI model that matches the performance of OpenAI’s o1 model in mathematics and coding tasks. Remarkably, the model was developed at a fraction of the cost—just $450—and in a mere 19 hours, setting a new benchmark for affordability in artificial intelligence development.
The NovaSky research team, behind the creation of the Sky-T1-32B-Preview model, demonstrated its ability to compete with OpenAI’s o1-preview. The project utilized only eight Nvidia H100 GPUs, significantly reducing expenses compared to traditional AI model training. The model leverages the capabilities of Qwen2.5-32-Instruct, an open-source framework, and was trained using data processed through another open-source model, QwQ-32B-Preview.
According to the research team, the development process involved integrating diverse datasets requiring complex reasoning. “We combined data from various fields that demand intricate problem-solving,” the team explained. “We then reprocessed the data using GPT-4o-mini to enhance its performance.”
In tests, the Sky-T1-32B-Preview model performed on par with or even surpassed OpenAI’s o1-preview in mathematics and coding benchmarks. However, it fell short in more advanced physics tests, such as the GPQA-Diamond benchmark. Despite this, NovaSky has made all components of the model—including data, weights, and technical details—open-source, fostering transparency and collaboration in the AI community.
This breakthrough highlights the potential for developing high-level reasoning capabilities at low cost and in a short timeframe. While OpenAI prepares to launch its next-generation o3 model, NovaSky’s achievement underscores the effectiveness of open-source solutions in advancing artificial intelligence.
The development not only opens new possibilities in AI research but also demonstrates that high-performance models can be created with limited resources. This milestone could pave the way for more accessible and cost-effective innovations in the field, democratizing AI development and encouraging broader participation.
