Shopee Built Custom LLM to Optimize E-commerce
The company deployed a 245-billion-parameter model to replace general-purpose AI, slashing production costs by 90%.
Updated on Sept. 24, 2026 in Artificial Intelligence
Live Poll
Do you believe companies building custom AI models will provide you with a better shopping experience?
Shopee has deployed Compass, a multilingual large language model trained on the NVIDIA AI Factory, to manage the majority of its e-commerce artificial intelligence traffic. The company reports the model now outperforms general-purpose alternatives on regional e-commerce tasks.
Why it matters
The shift was driven by the failure of off-the-shelf models to handle complex multilingual requirements across Shopee's Southeast Asian markets, which previously resulted in slower support resolution and missed fraud detection. By building a domain-specific model, the company has drastically improved operational efficiency for a platform facilitating over $136 billion in annual gross merchandise value.
Compass-v3 is a mixture-of-experts model—an architecture that activates only specific network paths per query—containing 245 billion total parameters with 71 billion active parameters. The model was trained on 12 trillion tokens of multilingual data.
The players
Shopee
An e-commerce platform facilitating over $136 billion in annual gross merchandise value across Southeast Asia, Taiwan, and Brazil.
NVIDIA
The semiconductor and computing company providing the AI Factory infrastructure used to train Shopee's multilingual models.
The details
The model training pipeline incorporates pretraining, instruction following, and agentic reinforcement learning—a method where the model learns to complete complex tasks by interacting with an environment to reach a goal. Shopee also implemented Multi-Token Prediction, which allows the model to predict several upcoming tokens simultaneously rather than one by one, increasing efficiency. To validate performance, researchers developed EcomEval, a custom benchmark designed to assess AI capabilities specifically for Southeast Asian retail datasets.
Timeline
Monthly API token usage surged from 3 billion to 340 billion over the past eight months.
The Tech Race
The move to custom, domain-specific models like Compass follows the industry trend of moving away from generalist AI toward specialized mixture-of-experts architectures. Shopee's development marks a departure from reliance on general-purpose model providers by internalizing the full training and deployment stack.
Users in Southeast Asia, Taiwan, and Brazil will see faster support resolution times and more effective automated fraud detection. The platform's 113x growth in API token usage suggests these capabilities are now integrated into the core retail experience for millions of shoppers.
The takeaway
Shopee's ability to reduce production costs by 90% while scaling to 340 billion monthly tokens demonstrates the massive efficiency gains available to companies that build domain-specific AI. Observers should track if other e-commerce giants release custom benchmarks comparable to EcomEval to validate their own proprietary model performance.
Further reading
For more on the development of specialized models, see the latest updates in Artificial Intelligence.
Source note: This article includes information reported by Frontier Enterprise.
Live Poll
Do you believe companies building custom AI models will provide you with a better shopping experience?
