AvitoTech/avision

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Oct 21, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

AvitoTech/avision is a 7 billion parameter Visual-Language Model (VLM) specifically adapted for the Russian language and the Avito domain. This model understands images and text together, excelling at tasks like describing photos, answering image-based questions, verifying description-photo consistency, and performing OCR for brands and text. It achieves a 50% speedup on Russian data and improved quality on Russian and domain-specific tests compared to its base model, Qwen2.5-VL-7B-Instruct.

Loading preview...

AvitoTech/avision: Russian-Optimized Visual-Language Model

AvitoTech/avision is a 7 billion parameter Visual-Language Model (VLM) developed by AvitoTech, specifically engineered for the Russian language and the Avito domain. It processes both visual and textual information to perform a variety of tasks, making it highly suitable for e-commerce and content moderation.

Key Capabilities and Optimizations

  • Multimodal Understanding: Describes images, answers questions based on visuals, verifies consistency between images and text, and extracts text (OCR) including brands and inscriptions.
  • Russian Language Focus: Adapted with a custom Russian tokenizer and fine-tuned on a large corpus of Russian text, resulting in a 50% speedup for Russian data processing compared to its base model, Qwen2.5-VL-7B-Instruct.
  • Domain-Specific Data: Trained on a proprietary Russian multimodal dataset, including ~200k ad images and ~1M question-answer pairs, supplemented by carefully localized open-source datasets.
  • Performance Improvements: Shows improved quality on Russian-specific benchmarks, including AvitoImageGen_RU (+6%), MMMU_RU (+1%), and RealWorldQA_RU (+1%), while maintaining universal VLM skills.
  • Efficiency: The model size is reduced to 7.4B parameters (from 8.3B), and its optimized tokenizer leads to fewer tokens for equivalent Russian examples, contributing to faster inference.

Use Cases within Avito

  • Automated Description Generation: Automatically creates product descriptions from photos.
  • Search Keyword Extraction: Extracts key features from images to enhance search functionality.
  • OCR and Normalization: Identifies and normalizes brands and text from images.
  • Streamlined Listing Creation: Facilitates one-click ad submission based on product photos.
  • Internal Tools: Supports internal annotation and moderation processes.

Finetunes
1 models