VK releases kazRush, a compact translation model for Russian and Kazakh
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A compact specialised model can improve translation for a low-resource language pair without requiring a multilingual system with billions of parameters.
The model has approximately 197 million parameters, was trained from scratch on a cleaned corpus of 9,054,960 sentence pairs and is available in both directions. In VK's comparison on FLORES+ devtest for Russian-to-Kazakh translation, kazRush scored 16.2 BLEU and 51.8 chrF, outperforming the tested NLLB variants on those metrics; its COMET score was 0.883 versus 0.889 for NLLB 3.3B. In the reverse direction, kazRush outperformed NLLB distilled 600M but trailed the larger NLLB models. At LoResMT 2026, the Russian-to-Kazakh version won its track with 76.24 chrF on the private test set. The work also showed that increasing the architecture to 738 million parameters did not by itself improve the result.
Sources 2