Researchers have unveiled Sky-T1, a ‘thinking’ AI model that can be trained for less than $450.

So-called artificial intelligence is becoming easier – and cheaper – to develop.

On Friday, NovaSky, a team of researchers from UC Berkeley’s Sky Computing Lab, released the Sky-T1-32B-Preview, a prototype that competes with an earlier version of OpenAI’s o1 on many benchmarks. The Sky-T1 looks to be the first open-source solution to be had repeated from the beginning; the group produced the data they used for training and the appropriate training plan.

“Incredibly, the Sky-T1-32B-Preview was trained for less than $450,” the team wrote in a statement. blog post“it shows that it is possible to imitate the higher faculties of thinking freely and effectively.”

$450 may not sound cheap. But it wasn’t long ago that the cost of training a model with similar behavior was high it was often worth millions of dollars. Artificial training data, or training data produced by other models, has helped to reduce costs. Palmyra X 004, a model has just been released by the AI ​​company Author, trained almost completely artificialit is estimated that it only cost $700,000 to complete.

Unlike most AI, reasoning models are self-diagnostic, which it helps them avoid some of the pitfalls that often attract moths. Deliberative models take longer – often seconds to minutes longer – to arrive at answers compared to a conventional non-deliberative model. The good thing is, they tend to be reliable in fields like physics, science, and math.

The NovaSky team says it used another model, Alibaba’s QwQ-32B-Previewto create the basics of Sky-T1 training, then “saved” the data fusion and added OpenAI. GPT-4o-mini to reorganize the data to make it more accessible. Training the 32 billion-parameter Sky-T1 took about 19 hours using a rack of 8 Nvidia H100 GPUs. (Parameters correspond to the model’s ability to solve problems.)

According to the NovaSky team, the Sky-T1 outperforms the preview version of the o1 on MATH500, a series of “competitive” math challenges. The model also beats the o1 benchmark for critical problems from LiveCodeBench, a code review.

However, the Sky-T1 falls short of the o1 on the GPQA-Diamond, which contains questions about science, biology, and chemistry that a PhD student would expect to know.

Also worth noting is OpenAI’s GA release of o1 is a more powerful model than the o1 model, and that OpenAI is expected to release better concepts, o3in the coming weeks.

But the NovaSky team says the Sky-T1 only marks the beginning of their journey to create open models with high-end concepts.

“Moving forward, we will focus on developing better models that retain critical thinking and exploring new techniques that improve the models’ efficiency and accuracy during testing,” the team wrote in a post. “Stay tuned as we move forward on this adventure.”