Recently, the Sky Computing Lab team at the University of California, Berkeley released Sky-T1-32B-Preview, an open-source inference artificial intelligence model that marks the development of inference AI easier and cheaper. The model performed well in multiple key benchmarks, even comparable to the earlier versions of OpenAI versions of o1.
The Sky-T1's training costs are impressive, at just $450, which means that the replication of high-level reasoning capabilities has become more affordable and efficient. While the $450 fee may not sound low, it is a huge drop compared to the training cost that cost hundreds of millions of dollars a few years ago. With the help of synthetic training data, i.e. training data generated by other models, the cost is significantly reduced. AI company Writer recently released the Palmyra X004, which relies almost entirely on synthetic data, and the development cost is only $700,000.
Unlike most AIs, inference models are able to self-check effectively, which makes them more reliable when dealing with some common problems. Inference models usually take more time when deriving solutions, which can take seconds to minutes, but in fields such as physics, science, and mathematics, the reliability advantages are significant.
The NovaSky team used another inference model - Alibaba's QwQ-32B-Preview to generate the initial training data of Sky-T1 and "planned" the data. Then, using OpenAI's GPT-4o-mini to re-do the data. Organize into a more operational format. It only takes about 19 hours to train Sky-T1 with 3.2 billion parameters, using a set of 8 Nvidia H100GPUs. The number of parameters is roughly related to the model's problem-solving ability.
According to the NovaSky team, Sky-T1 outperformed the early preview version of o1 on a MATH500 collection of “contest-level” mathematical challenges. In addition, Sky-T1 has encountered more difficulties in LiveCodeBench than the preview version of o1. However, Sky-T1 performed worse than the o1 preview in the GPQA-Diamond tests when it comes to issues involving physics, biology and chemistry.
It should be noted that OpenAI's GA version o1 is more powerful than the preview version, and OpenAI is expected to release an even better inference model o3 in the next few weeks. However, the NovaSky team said that Sky-T1 is just the beginning of their journey to develop an open source model with advanced reasoning capabilities.
“Looking forward, we will focus on developing more efficient models to maintain strong inference performance and explore advanced technologies that further enhance model efficiency and accuracy,” the team wrote in a blog. “Stay tuned for our progress on these exciting projects.”