Multilingual Large Language Models (LLMs) often exhibit degraded performance for languages other than English due to the imbalance in their training data. Directly adapting these models to a new language, such as Russian, carries the risk of catastrophic forgetting of their original capabilities and demands significant computational resources. The article introduces Ruadapt: a comprehensive and computationally efficient methodology for language adaptation of LLMs, featuring tokenizer replacement. A full adaptation of a single Qwen3-8B model version with our methodology requires less than 2000 GPU h, while subsequent adaptations of other versions are up to ten times less resource-intensive due to the modular nature of the procedure’s steps. An optimal configuration achieves up to an 80% speed-up in generation, with full preservation of long-context capabilities and only minor degradation in instruction-following performance. The authors conduct a detailed empirical study of each adaptation step to identify optimal hyperparameters and to assess the impact of each key stage on the final quality. These resulting guidelines are implemented in the current generation of Ruadapt models, such as RuadaptQwen3-32B-Hybrid. We are open-sourcing our models, code, and datasets to provide the research community with a validated and cost-effective strategy for developing high-quality, language-specific models.
Tikhomirov et al. (Fri,) studied this question.