Reasoning AI is moving into a new phase: not just more capable, but easier to build and share. NovaSky, a team of researchers based out of UC Berkeley's Sky Computing Lab, has released Sky-T1-32B-Preview, a reasoning model that can be trained at a cost that would have sounded improbable not long ago.
The release matters because Sky-T1 appears to be the first truly open source reasoning model in the sense that it can be replicated from scratch. NovaSky released both the dataset used to train it and the training code needed to reproduce the work.
Why Sky-T1 Stands Out
Sky-T1-32B-Preview is designed as a reasoning model, meaning it is built to work through problems in a more deliberate way than a typical nonreasoning AI model. According to the source, reasoning models effectively fact-check themselves, which can help them avoid some common pitfalls that normally trip up AI systems.
That extra checking comes with a trade-off. Reasoning models usually take longer to answer, often seconds to minutes longer than a typical nonreasoning model. The benefit is that they tend to be more reliable in areas such as physics, science, and mathematics.
NovaSky says Sky-T1-32B-Preview was trained for less than $450. The team framed that cost as evidence that high-level reasoning capabilities can be replicated affordably and efficiently.
"Remarkably, Sky-T1-32B-Preview was trained for less than $450,"
The figure is striking because the source notes that training a model with comparable performance often ranged in the millions of dollars not long ago. Sky-T1 is therefore part of a broader shift: advanced AI capabilities are no longer only defined by very large training budgets.
The Role of Synthetic Training Data
One reason costs are falling is synthetic training data, which means training data generated by other models. The source points to synthetic data as a major factor in making model development cheaper.
NovaSky used Alibaba's QwQ-32B-Preview, another reasoning model, to generate the initial training data for Sky-T1. The team then curated the data mixture and used OpenAI's GPT-4o-mini to refactor the data into a more workable format.
This workflow shows how model development can build on existing AI systems. Instead of relying only on manually produced data, researchers can use other models to create or reshape training material, then refine that material for a specific goal.
The source also mentions Palmyra X 004, a model recently released by AI company Writer. It trained almost entirely on synthetic data and reportedly cost just $700,000 to develop. That is still far above Sky-T1's reported training cost, but it points in the same direction: synthetic data is changing the economics of AI training.
How Sky-T1 Was Trained
Sky-T1 is a 32-billion-parameter model. The source explains that parameters roughly correspond to a model's problem-solving skills.
Training took about 19 hours using a rack of 8 Nvidia H100 GPUs. That combination of a defined model size, a short training window, released training code, and a released dataset is central to the open source claim around Sky-T1.
The important distinction is not simply that the model is available. The source says Sky-T1 appears to be truly open source because it can be replicated from scratch. For researchers, developers, and organizations interested in reasoning AI, reproducibility is a major part of the value.
That does not mean Sky-T1 is the strongest reasoning model available. It means NovaSky has shown a path for creating a competitive reasoning model with public ingredients and a comparatively low training bill.
Where It Beats o1 Preview, and Where It Falls Short
NovaSky says Sky-T1 performs better than an early preview version of OpenAI's o1 on MATH500, a collection of "competition-level" math challenges. It also beats the preview of o1 on difficult problems from LiveCodeBench, a coding evaluation.
Those results are meaningful because math and coding tasks are natural proving grounds for reasoning models. They test whether a system can follow multi-step logic, not just produce fluent text.
But the comparison is not one-sided. Sky-T1 falls short of the o1 preview on GPQA-Diamond, which contains physics, biology, and chemistry-related questions a PhD graduate would be expected to know.
There is another important limit to the comparison. The source notes that OpenAI's GA release of o1 is stronger than the preview version of o1. OpenAI is also expected to release an even better-performing reasoning model, o3, in the weeks ahead.
In other words, Sky-T1 is not presented as a definitive leader over OpenAI's latest reasoning systems. Its significance is that an open source model, trained for less than $450, can compete with an earlier o1 preview on key benchmarks.
What Comes Next for Open Source Reasoning AI
NovaSky describes Sky-T1 as a beginning, not an endpoint. The team says it plans to continue developing open source models with advanced reasoning capabilities.
The next focus is efficiency and accuracy. NovaSky says it will work on more efficient models that maintain strong reasoning performance, while also exploring advanced techniques that improve efficiency and accuracy at test time.
That direction is logical given the model's current profile. Sky-T1 already shows that reasoning AI can be made cheaper and more reproducible. The next challenge is keeping that accessibility while improving performance across more demanding evaluations.
For the wider AI field, the release adds pressure to a fast-changing question: who gets to build advanced reasoning systems? If models like Sky-T1 can be replicated from scratch, the answer may increasingly include researchers and teams far beyond the largest AI labs.