Sky-T1-32B-Preview: How a $450 Model Could Change AI Forever

2 min read
Co-founder of JustAINews
Share
Key Points
  • Sky-T1-32B-Preview was trained under $450 using Berkeley’s Sky Computing Lab resources, with support from Lambda Labs and Anyscale.
  • Open-source collaboration fosters advanced reasoning, enabling researchers worldwide to replicate and refine these efficient AI models.
  • Combining math and coding data was critical for maintaining robust logical capabilities across both challenging domains.
AI Development
Credits: Markus Spiske / Pexels

NovaSky has presented Sky-T1-32B-Preview, an open-source reasoning model trained for under $450 but strong enough to rival o1-preview on leading math and coding benchmarks. Developed at UC Berkeley’s Sky Computing Lab with the support of Lambda Labs and Anyscale, it demonstrates that high-level reasoning models no longer need extravagant budgets.

"By openly sharing all code, data, and weights, NovaSky offers full transparency, enabling researchers, developers, and learners to recreate or advance the project. Remarkably, Sky-T1-32B-Preview was trained for less than $450, demonstrating that it is possible to replicate high-level reasoning capabilities affordably and efficiently. All code is open-source."

Blog Post, NovaSky

The Importance of openness

A common perception is that large-scale AI demands costly training and closed-door methods. Sky-T1-32B-Preview challenges that idea by employing a balanced data mixture of math and coding problems, along with a careful sampling process to remove incorrect or low-quality outputs.

The result is a model that performs effectively across diverse tasks.

For the public, this openness means the model’s capabilities and limitations can be studied and improved in a collaborative environment.

How did they accomplish this?

NovaSky generated the initial training data using another open-source model with robust reasoning features. They then refined and reformatted the data with GPT-4o-mini, ensuring well-structured samples. Rejection sampling was used to filter out errors. Ultimately, they assembled 17K carefully vetted data entries drawn from math competitions, coding challenges, and puzzle sets.

In benchmark tests, Sky-T1-32B-Preview performed notably well on advanced math problems, including AIME24, and handled various coding tasks in LiveCodeBench with consistent accuracy.

These outcomes highlight a model that can address both specialized math queries and practical coding scenarios.

A shared vision for the future

Sky-T1-32B-Preview is just the starting line. NovaSky aims to refine and expand the model’s reasoning capabilities by exploring advanced techniques for improved efficiency at test time. They also plan to incorporate emerging HPC solutions to drive down costs even further.

While there’s excitement about how this open-source model compares to proprietary alternatives, NovaSky stands by its mission: anyone should be able to replicate and extend top-tier AI research.

By offering full transparency—data, code, model weights—they invite global collaboration. In turn, academics, students, and industry professionals can collectively propel the field forward, each group injecting its unique expertise to refine these powerful models.

Subscribe to JustAINews!

Get the industry's biggest AI news straight to your inbox.
Subscribe

Related posts

© Copyright 2025 - Just AI News - All Rights Reserved
linkedin facebook pinterest youtube rss twitter instagram facebook-blank rss-blank linkedin-blank pinterest youtube twitter instagram