Vikhr-Nemo-12B

Vikhr-Nemo-12B is a 12-billion-parameter instruction-tuned language model developed by VikhrModels, built on top of Mistral-Nemo-Instruct-2407. It was adapted specifically for high-quality generation in Russian and English, with a 128k-token context window inherited from its base model.

The model was trained with a multi-stage recipe: supervised fine-tuning on a 150k-example synthetic instruction dataset, followed by alignment with SMPO, a DPO variant developed for the project. It also ships a Grounded RAG mode that identifies document identifiers for grounded answers. The ACL paper on the Vikhr family notes the team adds no restrictions beyond the base model's license.