> ## Content Index
> Fetch the complete content index at: https://nordisk-ai.ghost.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# Sky-T1: Record-Cheap Model to Train
- URL: https://nordisk-ai.ghost.io/sky-t1-record-cheap-model-to-train/
- Published: 2025-02-10T12:58:11.000Z
- Updated: 2026-08-27T15:08:24.000Z
- Description: Reasoning-based AI models are changing the game - and now they can be trained for under $450. How will Sky-T1 and open source change the future of AI?
- Author: Anders Hagen
- Tags: AI Models, AI Research, AI training, AI Development, aisweden, article, GPU, cost-effective AI, nordiskai, novaSky, open source, reasoning-ai, sky-t1, synthetic data, techcrunch, uc berkeley, #Migrated-1787841634512, #wp, #wp-post, #Import 2026-08-27 14:41, #Import 2026-09-07 13:41

Researchers from UC Berkeley's Sky Computing Lab have presented Sky-T1-32B-Preview, a reasoning model that competes with earlier versions of OpenAI's o1 in several key areas. Sky-T1 is groundbreaking as a truly open-source model, with both the dataset and training code made available to the public. The model was trained for less than $450, showing how advanced AI can be developed cost-effectively.

### A Price Revolution in AI Training

Just a few years ago, the cost of training AI models of the same caliber could run into the millions of dollars. Sky-T1's development is an example of how synthetic training data, generated by other AI models, has drastically reduced costs. For example, the model Palmyra X 004 was developed almost entirely on synthetic data for about $700,000\. Sky-T1 demonstrates that high-performing reasoning models can now be created for a fraction of that sum.

### What Sets Reasoning Models Apart From Other AI Models?

Reasoning models have a unique ability to "fact-check" themselves and avoid many common pitfalls other models tend to fall into. They often take longer to solve problems, from seconds to minutes, but the result is higher reliability in areas like physics, science, and math. For Sky-T1, a mix of data from Alibaba's reasoning model QwQ-32B-Preview was used and refactored with the help of OpenAI's GPT-4o-mini to create a more useful dataset. The model, which has 32 billion parameters, was trained for 19 hours on eight Nvidia H100 GPUs.

### Strengths and Weaknesses of Sky-T1

On tests like MATH500, a collection of advanced math problems, Sky-T1 outperforms an earlier version of OpenAI's o1\. It also performs strongly on LiveCodeBench, a code evaluation platform. However, it lags slightly behind on GPQA-Diamond, a test focused on graduate-level questions in physics, biology, and chemistry.

It's worth noting that OpenAI's later versions of o1, as well as upcoming models like o3, are expected to be even more powerful.

### A New Era of Open-Source AI

The NovaSky team sees Sky-T1 as the start of a journey toward developing more efficient open-source models with advanced reasoning capabilities. Their goal is to improve both performance and resource efficiency. This initiative could change the landscape for how AI models are trained and used in the future.