Avito: modelos próprios A-Vibe e VLM automatizam a moderação e vendas em mais de 200 milhões de anúncios
Over 98% of all marketplace listings are now moderated automatically without human intervention, while contextual chat analysis significantly reduced multi-step phishing incidents.
| Company & Sector | Авито (Avito) — E-commerce e classificados |
|---|---|
| Models & AI Stack | A-Vibe LLM (8B) · Custom RU Tokenizer · Vision-Language Models (VLM) · Graph Neural Networks · Triton / vLLM · Real-Time Antifraud |
| Cronologia | 2023 — scale-up of ML anti-fraud and CV photo verification; 2024 — release of proprietary A-Vibe LLM with custom Russian tokenizer and VLM photo-to-listing autofill; 2025–2026 — 20+ production GenAI use cases across search, chats, and support. |
| Verification | Verificado 2026-10 · 3 primary sources |
Contexto e escala do negócio
Avito is Russia's largest classifieds marketplace and one of the highest-traffic commerce platforms globally, hosting over 200 million active listings across Goods, Auto, Real Estate, Services, and Jobs with more than 60 million monthly active users.
At this scale, legacy rule engines and narrow classifiers hit diminishing returns: sellers upload sparse descriptions, take ambiguous photos, and fraudsters continuously mutate social-engineering patterns inside buyer-seller chats.
Between 2024 and 2025, Avito developed its own family of foundation models—including the A-Vibe LLM with a custom Russian-optimized tokenizer and multimodal Vision-Language Models (VLMs)—backing an announced 12 billion ruble GenAI investment program targeting over 21 billion rubles in cumulative business return.
Problema e gargalos operacionais
The first bottleneck in any classifieds marketplace is listing creation friction. Casual sellers dislike spending 15 minutes filling out dozens of structured attributes, resulting in poorly indexed listings and repetitive buyer questions.
The second challenge is real-time moderation and anti-fraud. Every second, the platform must inspect hundreds of newly uploaded photos, descriptions, and thousands of chat messages for prohibited items, phishing attempts, and duplicate spam without delaying legitimate sellers.
Third, inference economics: running unoptimized 70B+ parameter models with bloated Cyrillic tokenization across tens of millions of daily interactions would break unit economics.
Arquitetura e implementação de IA
Avito's ML organization built compact domain-adapted models (the A-Vibe LLM family with a custom Russian e-commerce tokenizer) alongside specialized Vision-Language Models (VLMs). Optimizing the tokenizer for Russian commerce vocabulary boosted token throughput by over 40% while slashing GPU memory footprint.
During listing creation, a seller simply uploads photos: the VLM extracts category, brand, model, color, and condition, populates structured filters, and drafts a clear description—cutting listing creation time nearly in half.
In trust & safety, an ensemble of CV, graph neural networks, and LLMs resolves over 98% of moderation decisions automatically in sub-second latency, while chat AI assistants help buyers compare real estate or auto listings and automate first-line replies for professional merchants.
“"A custom tokenizer and compact domain-specific models allow us to serve generative AI to tens of millions of daily users with sub-second latency and sustainable unit economics." — AvitoTech Engineering
Impacto mensurável e economia
Over 98% of all marketplace listings are now moderated automatically without human intervention, while contextual chat analysis significantly reduced multi-step phishing incidents.
Photo-to-listing VLM autofill cut listing creation time by 40–50% and lifted buyer contact conversion by ensuring structured search attributes are completely populated.
Across more than 20 production GenAI use cases spanning search, trust & safety, customer support, and merchant tools, Avito projects over 21 billion rubles in annualized economic impact by 2028.
Lições para a indústria
At tens of millions of DAU, an 8B domain model with a custom Cyrillic tokenizer beats giant general LLMs on latency and cost.
Extracting structured catalog attributes from photos via VLM drives more GMV than prose generation alone because search filters depend on structured fields.
Marketplace trust & safety requires fusing text, image embeddings, and account graph signals in real time.
Publishing technical benchmarks and open-weights artifacts (AvitoTech) significantly accelerates senior ML hiring.
Fontes primárias verificadas
Perguntas frequentes sobre a notícia
Why did Avito train its own A-Vibe model and tokenizer instead of using stock open-weight models?
Default multilingual tokenizers fragment Russian words and e-commerce terms inefficiently. Rebuilding the vocabulary and continued pretraining cut token counts and accelerated inference by 40%+.
How does the VLM assist sellers when posting an item?
From uploaded photos alone, the VLM detects item category, brand, color, and condition, pre-filling structured filter fields and drafting a listing description.
Does automated moderation replace human reviewers completely?
Over 98% of routine listings are cleared or blocked automatically in seconds, while ambiguous edge cases and appeals are routed to human specialists with pre-highlighted risk signals.
Executive AI Consulting, Custom Agents & Team Workshops
Work directly with Zhemal Khamidun at hamidun.com — enterprise AI architecture, production RAG & multi-agent systems, and executive training.
O essencial da IA — uma vez por semana
Sete histórias que realmente importaram, escolhidas a dedo. Sem ruído nem releases.
Pronto! Verifique seu e-mail para a confirmação.