MetalGPT-1

2025: Product Announcement

In December 2025, Nornickel introduced MetalGPT-1, its own domain language model for metallurgy and mining. The model became the first in the family of large language models of the company. The company announced the launch of the model on December 9, 2025.

According to the press service of Norilsk Nickel, the developed model forms a single language layer for engineering, technological, production and corporate tasks. On its basis, the company creates personal AI assistants and autonomous agents that are introduced into operational processes. Using a domain model reduces hallucinations and improves the quality of decisions made based on artificial intelligence recommendations.

𝓲

Фото: Freepik

The language model includes 32 billion parameters and is trained on 10 gigabytes of specialized texts on metallurgy and mining. This volume is comparable to half of the English-language Wikipedia. The key competitive advantage of the model is its unique data quality. The training was conducted on more than 1 million documents that are not available in open sources.

Training data includes technological protocols, internal regulations and instructions of enterprises, design and construction documentation, patents, R&D reports, scientific and technical literature. All data has undergone multi-stage cleaning and anonymization, which allowed the use of industry knowledge without disclosing trade secrets.

In addition, about 500 thousand question-response and instructional pairs were created on the basis of real production and scientific tasks. This helps the model better capture causal relationships in technological processes and produce error-resistant responses. Synthetic data teaches the model to understand the context of specific production situations.

The development of the MetalGPT-1 took about a year. It took six months to collect and prepare data, two months for basic training and two more for domain adaptation and fine-tuning the model. The long stage of data preparation reflects the complexity of working with industrial documentation and the need to structure it.^[1]

Notes