A-vision

A-Vision is a Russian-language vision-language model (VLM) built upon the Qwen2.5-VL-7B-Instruct architecture. Its uniqueness lies in its targeted domain adaptation for understanding images within the context of the Russian language and marketplace-specific nuances. Like A-Vibe, it features an optimized tokenizer, which reduced its size to 7.4B parameters and increased processing speed for Russian multimodal queries by 50%. However, its key differentiator is a proprietary training corpus collected by Avito: approximately 200,000 product images paired with nearly one million Russian-language image–question–answer triplets. This enables the model to deeply comprehend the context of everyday objects and classified ads.

Technically, A-Vision underwent a multi-stage adaptation process: first, the tokenizer was replaced with a Russian-optimized version, followed by a freeze→unfreeze training protocol applied to the LLM component using a large Russian text corpus. Next, multimodal supervised fine-tuning (SFT) was performed on Avito’s custom dataset, and finally, Direct Preference Optimization (DPO) was applied to ensure response safety.

The model demonstrates balanced expertise. It leads on Avito’s key business benchmark, AvitoImageGen_RU (0.7668), which evaluates the quality of automatic product description generation from images, delivering a +6% improvement over the base model. At the same time, it preserves—and even enhances—the original model’s strengths in English-language tasks, outperforming Qwen2.5-VL-7B-Instruct on the DocVQA_EN test (94.97), which assesses the ability to extract information from documents and forms.

A-Vision’s use cases are directly tied to its multimodal nature and tailored training. It is designed to automate visual content processing: instantly generating product titles and descriptions from photos, extracting text, brands, and attributes from images (OCR), verifying image–description alignment, and assisting with visual content moderation on platforms.

Announce Date: 21.10.2025
Parameters: 7B
Context: 128K
Layers: 28
Attention Type: Full Attention
VRAM requirements: 12.6 GB using 4 bits quantization
Developer: AvitoTech
Transformers Version: 4.57.1
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore A-vision 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.