A-vibe

A-vibe is a Russian-language large language model developed by Avito, built upon the open-source Qwen3-8B-Base. Its key innovation lies in a unique approach to Russian language adaptation: rather than merely fine-tuning the model, the developers completely replaced the tokenizer, merging English tokens from the original Qwen3 with Russian tokens from a specially trained tokenizer. This hybrid approach achieves high tokenization efficiency for Russian text—using on average 22% fewer tokens for the same content—significantly accelerating inference and reducing the model size to 7.9 billion parameters. As a result, A-vibe processes Russian-language queries 15–25% faster than the base version.

Technically, A-vibe’s training pipeline comprised several critical stages: first, tokenizer adaptation on a corpus of 150 billion tokens (31% Russian and 31% English); next, supervised fine-tuning (SFT) on over 800,000 examples, including synthetic dialogues with function calling. This was followed by GRPO (Generalized Reinforcement Learning with Policy Optimization) to enhance mathematical reasoning and function-calling capabilities, and DPO (Direct Preference Optimization) to improve dialogue safety and quality. Special attention was paid to partially freezing embeddings during tokenizer adaptation—an innovative gradient-hooking technique that preserved the quality of representations for English tokens.

A-vibe demonstrates outstanding performance on Russian-language benchmarks: it outperforms the base Qwen3-8B on math_500_ru (68.6% vs. 54.6%). On the BFCL V3 function-calling benchmark, the model achieves 58.63%, confirming its strong function-calling capabilities. Most impressively, in the RU_ARENA ranking, A-vibe surpasses not only Qwen3-8B but also other Russian-language models significantly larger in size.

Use cases for A-vibe naturally stem from its architecture and strengths. It is ideally suited for building intelligent Russian-language chatbots and assistants, analyzing and summarizing text (including user inquiries and documents), generating and explaining code, and solving logical and computational tasks within educational, analytical, and service-oriented products.

Announce Date: 20.10.2025
Parameters: 7.9B
Context: 32K
Layers: 36
Attention Type: Full Attention
VRAM requirements: 10.7 GB using 4 bits quantization
Developer: AvitoTech
Transformers Version: 4.52.0.dev0
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore A-vibe capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.