According to the press service, these are the first Russian open-source models trained on e-commerce data and adapted for the Russian language. The company invested about half a billion rubles in their development.
The models are built on the base of Qwen 3 and Qwen 2.5-VL, include a proprietary tokenizer for the Russian language (doubling text generation speed), and support function calling - the ability to connect external tools to create AI agents. To evaluate Russian-language models, the Avito team adapted four international benchmarks - MATH-500, GPQA Diamond, DROP_RU, and BFCL V3 - and made them publicly available. Avito noted:
In MERA and RU Arena Hard tests, the A-Vibe model showed the best results in its class, outperforming larger neural networks such as GPT-4o mini, Gemma 3 27B, Claude 3.5 Haiku, and others.