AvitoTech

AvitoTech (Avito Tech LLC) is the technology division of the Avito group of companies, the largest classifieds platform in Russia with over 113 million monthly users. The company is registered in Moscow (7 Lesnaya Street) and operates as an independent legal entity, engaged in the development of software and databases. The organization has evolved from a team supporting a high-load classifieds platform to creating its own AI solutions, managing an infrastructure of over 3000 microservices with a team of 2700+ engineers.

Historically, AvitoTech has developed around e-commerce tasks: content moderation, search, recommendations, and computer vision. In 2024–2025, the company made a strategic shift towards developing its own large language models.

AvitoTech developed the LEP Initialization (Language-Specific Embedding Projection) methodology for the smart initialization of embeddings for new tokens. The company implemented the SFT-mixed Tokenizer Training technique, where an SFT dataset is mixed directly into the tokenizer training process to balance token representation across code, text, and e-commerce-specific data. Based on this research, in the fall of 2025, AvitoTech open-sourced two model families — A-Vibe and A-Vision — specialized for the Russian language and e-commerce. Both models were released under the open-source Apache 2.0 license on Hugging Face, allowing researchers and developers in the Russian-speaking segment to use them without restrictions for commercial and research purposes.

Related models