Reported improvements include higher benchmark scores, expanded token usage, lower fraud-processing costs and broader enterprise deployment

Each month, e-commerce platform Shopee processes billions of queries across South-east Asia, Taiwan and Brazil, while supporting hundreds of millions of monthly active users and more than US$136bn in annual gross merchandise value.

The platform’s use of AI spans search, recommendations, fraud detection, logistics and customer service — but general-purpose language models had been performing inconsistently on the multilingual and domain-specific tasks involved. Shortcomings had been identified in areas such as Indonesian shopping queries, Thai product categorization and Vietnamese customer-intent analysis.

Shopee noted that some open-source models scored below 52% on selected multilingual e-commerce tasks. To assess the problem, its researchers developed a benchmark covering five regional ecommerce categories. They subsequently developed their own domain-specific language models and scaled their use in production. According to the firm, monthly API-token consumption then increased from 3bn to 340bn over eight months. The reported implementation included:

  • A mixture-of-experts architecture with 245 billion total parameters and 71 billion active parameters per token
  • Pre-training on 12 trillion curated multilingual and ecommerce-related tokens, including Indonesian, Thai, Filipino, Malay, Tagalog and Portuguese data
  • Proprietary training data derived from product titles, reviews and ratings, supplemented by synthetic instruction data
  • An 80,000-token byte-pair encoding vocabulary and curriculum learning that progressively increased sequence length and the proportion of lower-resource languages
  • Preference optimization that weighted training tokens according to semantic importance, followed by instruction tuning, preference alignment and reinforcement learning
  • Distributed training and inference across thousands of data-centre GPUs, with software for large-scale model training, post-training and inference optimization
  • Deployment in search, recommendations, fraud detection, parcel recovery and multilingual customer service

The platform’s latest model scored 94% on e-commerce question answering, compared with 85% for GPT-4o; 92% on shopping concepts, compared with 82%; and 84% on user understanding, compared with 76%. Fraud-detection workflows were now 50 times more efficient than manual review, and reduced processing costs by 90%, according to its announcements.

Jason Dong Jian, Director and AI Platform Lead, Sea Limited, said the increase in monthly tokens “validated our ability to operationalize AI at platform scale” and supported wider deployment across the business.

Nvidia, the firm that consulted on Shopee’s AI improvements, said its GPU infrastructure and software supported the latter firm’s model-development and inference workloads, including distributed pre-training, model alignment and production serving.