Nornickel Releases MetalGPT-1 - 32B LLM Trained on a Million Closed Technological Documents in Metallurgy and Mining

Nornickel has released MetalGPT-1 - a 32B LLM trained on a million closed technological documents in metallurgy and mining. It is important not only that this is a domain-specific model but also the type of data it was fed. Technological protocols, regulations, R&D, construction and project documentation are not texts in the conventional ML sense. These are formalized fragments of the production world: the language of processes, chains, constraints, and risks. By training the LLM on such a corpus, the company effectively creates a separate 'data-reality layer' that universal models simply do not see. This is the main news. A new paradigm emerges: not 'adapting GPT to the domain' but 'building AI around the industrial world as the primary source of data.' MetalGPT-1 is just the first example. There will be models for chemical engineering, logistics, energy, and construction. Each industry has its own language, its own dataset, its own reality. This means that domain-first LLMs are no longer an experiment; they are becoming infrastructure.