01 What happened
Shopee, the leading ecommerce platform in Southeast Asia, Taiwan, and Brazil, has developed its own large language model Compass v3, a 245-billion parameter mixture-of-experts model tailored for multilingual ecommerce. According to an NVIDIA case study, the model now handles the majority of Shopee’s AI traffic, with monthly API tokens growing from 3 billion to 340 billion in eight months – a 113x increase.
02 Key details
- Compass v3.5 surpasses GPT-4o, GPT-4.1, Qwen3-235B, and DeepSeek v3.1 across all five EcomEval task categories, achieving 94% on Ecom QA compared to GPT-4o’s 85%.
- Monthly API tokens grew from 3 billion to 340 billion (113x) in eight months, driven by NVIDIA A100, H100, and Blackwell GPUs.
- Fraud detection powered by NVIDIA TensorRT-LLM demonstrated a 50x faster efficiency over manual review with 90% lower processing costs in production.
- The reasoning-focused Compassmax v3.5 variant attains 93 scores on Arena Hard v2 and 83 on AIME25, matching or exceeding proprietary reasoning models like GPT-5-Thinking.
- Shopee's Compass model handles the majority of its AI traffic, supporting expansion to additional languages, markets, and use cases.
03 Why it matters
General-purpose LLMs were insufficient for specialized, multilingual e-commerce tasks, prompting Shopee to build a vertical-domain model using NVIDIA’s AI factory stack, including Megatron-Core and NeMo.
04 Who it matters to
E-commerce platforms, AI infrastructure engineers and multilingual NLP researchers.
Original source