NovaSky, a research team at UC Berkeley’s Sky Computing Lab, has launched Sky-T1-32B-Preview, a groundbreaking reasoning AI model. This release is significant not only for its open-source status but also for its affordability, with training costs under $450.

The Sky-T1 model, trained in just 19 hours using 8 Nvidia H100 GPUs, showcases the potential of efficient AI development. By using synthetic data generated by Alibaba’s QwQ-32B-Preview and OpenAI’s GPT-4o-mini for data refinement, NovaSky reduced the cost and time of training significantly.

Reasoning AI models differ from traditional AI by their self-checking capabilities, which enhance their reliability in solving complex problems. While they take longer to process tasks, their accuracy makes them valuable in fields like mathematics, physics, and coding.

Sky-T1 performed strongly on MATH500, a benchmark of advanced math problems, and LiveCodeBench, a coding test. It surpassed OpenAI’s o1 preview model in these domains. However, it lagged behind on GPQA-Diamond, which tests scientific reasoning in fields like biology and physics.

Despite these limitations, Sky-T1 represents a major step in democratizing advanced AI. Historically, training models with similar performance cost millions of dollars. NovaSky’s approach proves high-level reasoning AI can be developed at a fraction of that cost.

OpenAI’s o1 GA release and its upcoming o3 model still lead in overall performance. Yet, Sky-T1's open-source nature and affordability mark it as a significant development.

The NovaSky team has stated their commitment to advancing this technology. "We aim to develop more efficient models while enhancing reasoning performance," they wrote in a blog post.

SOURCE: TECHCRUNCH