Researchers open source Sky-T1, a 'reasoning' AI model that can be trained for less than $450

Researchers from UC Berkeley’s Sky Computing Lab have unveiled Sky-T1-32B-Preview, a new open source “reasoning model” that rivals an earlier version of OpenAI’s o1 on math- and coding-focused benchmarks. Significantly, Sky-T1 reportedly cost less than $450 to train—a stark contrast to the millions of dollars often spent on comparable large-scale models not long ago. The team credits synthetic training data for helping cut down expenses; Sky-T1’s initial dataset was generated by Alibaba’s QwQ-32B-Preview, and then refined with OpenAI’s GPT-4o-mini. Training the 32-billion-parameter Sky-T1 took only 19 hours on a cluster of eight Nvidia H100 GPUs, illustrating both the growing affordability and rapid development pace for advanced AI projects.

Though Sky-T1 outperforms the preview version of o1 on complex math and coding tests, it lags behind o1 in a separate domain of graduate-level science questions, and also trails OpenAI’s more polished GA release of o1. Nonetheless, Sky-T1 stands as a promising step toward fully open source models with robust reasoning capabilities—models that can “fact-check” themselves and potentially avoid errors that often plague conventional AI. The Sky Computing Lab team plans to keep refining their approach, working on ways to boost both efficiency and performance for their future reasoning models.