Sky-T1: Train your own O1 preview model within $450

We introduce Sky-T1-32B-Preview, our reasoning model that performs on par with o1-preview on popular reasoning and coding benchmarks. 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.

Models such as o1 and Gemini 2.0 flash thinking that excel in reasoning have shown to solve complex tasks by producing a long internal chain of thought, among other advancements. However, the technical details and model weights are un-accessible, presenting a barrier to the participation of the academic and open-source communities.

In response, a few notable efforts have emerged to train open-weight reasoning models in the math domain, such as Still-2 and Journey. Concurrently, we, the NovaSky team at UC Berkeley, have been exploring various techniques to evolve the reasoning capabilities of base and instruct-tuned models. In this work, we achieve competitive reasoning performance not just in math, but also in coding in the same model.

To ensure our work benefits the broader community, we are fully committed to open-source collaboration. We open-source all details (i.e., data, codes, model weights) to enable the community to replicate and improve on our results easily:

Leave a Comment