Moscow, Russia — December 9, 2025 — Nornickel has unveiled MetalGPT-1, its proprietary domain-specific language model for the metallurgy and mining industry. The model is the first in the company’s family of large language models to be released as open source. Unlike universal models trained on general internet data, MetalGPT-1 is designed from the ground up to handle professional terminology, abbreviations, and complex technological chains, reducing the rate of hallucinations and improving the quality of decisions made based on AI recommendations.

The developed model forms a unified linguistic layer for engineering, technological, production, and corporate tasks. Nornickel is using it to build personal AI assistants and autonomous agents that are being integrated into the company’s operational processes.

The language model features 32 billion parameters and was trained on 10 gigabytes of specialized texts on metallurgy and mining—a volume comparable to half of the English-language Wikipedia. The model’s key competitive advantage lies in the unique quality of its data: training was conducted on over one million documents unavailable in open sources. These include technological protocols, internal regulations and enterprise instructions, design and construction documentation, patents, R&D reports, and scientific-technical literature. All data underwent multi-stage cleaning and anonymization, allowing for the use of industry knowledge without compromising trade secrets. Additionally, approximately 500,000 question-answer and instruction pairs were created based on real production and scientific tasks to help the model better grasp cause-and-effect relationships in technological processes and provide error-resistant answers.

“Metallurgy is one of the most complex industry domains with its own language of processes, abbreviations, and terms. Universal models trained on general web corpora lose accuracy with such specialized language, while large-scale models require colossal computing resources. MetalGPT-1 changes the rules of the game: it is the world’s first domain model with 32 billion parameters specifically optimized for metallurgy. On an industrial benchmark, it demonstrates the industry’s best level of understanding of metallurgical language using resources accessible for real industrial application. Every company can now adapt the model to its own tasks,” noted Danil Ivashechkin, Head of AI Development at Nornickel.

The development of MetalGPT-1 took approximately one year: six months were spent on data collection and preparation, two months on base training, and another two on domain adaptation and fine-tuning. For an objective quality assessment, the Nornickel team created an industrial benchmark for metallurgy—a set of question-answer pairs covering various processes in the mining and metallurgical industry—on which MetalGPT-1 consistently outperforms open universal models.

Nornickel has published the MetalGPT-1 model and the industrial benchmark on the Hugging Face platform, providing the industry with tools to develop specialized solutions and expand the ecosystem of industrial applications based on domain-specific language models.