Vikhr

Vikhr is a bilingual open-source instruction-following LLM built specifically for Russian by Aleksandr Nikolich and collaborators (ITMO, HSE, Yandex, MBZUAI). Unlike Russian adaptations that only add LoRA adapters, Vikhr adapts the tokenizer vocabulary for Russian and then continually pretrain and instruction-tune all weights, which improves both generation quality and token efficiency in Russian. Vikhr-5.2, built on Mistral-7B, is a representative release.

The team reports state-of-the-art results among open Russian models on Russian benchmarks, in some cases ahead of proprietary systems. Note for the record: one conference version of the Vikhr paper was retracted from the ACL Anthology, while the arXiv versions remain the canonical reference. Weights, instruction data, and code are publicly available.