Shopee’s LLMs Deliver Frontier-Level Performance With the NVIDIA AI Factory

Shopee’s AI team pretrained Compass-v3, a vertical-domain mixture-of-experts model with 245B total parameters and 71B active per token, designed for Southeast Asian ecommerce. It was trained on 12T tokens of curated multilingual corpora, including Indonesian, Thai, Filipino, Malay, Tagalog, and Portuguese, as well as large-scale synthetic ecommerce instructions using a mixed-training strategy.

To enhance alignment, Shopee proposed Optimal-Transport Direct Preference Optimization, or OTPO, drawn from seven cleaned and deduplicated source types and tokenized with a custom 80K BPE vocabulary, using AdamW and BFloat16 with a two-axis curriculum learning schedule that gradually increases sequence length and the share of low-resource Southeast Asian languages.

The compute foundation spans thousands of NVIDIA GPUs—including NVIDIA A100, NVIDIA H100, and NVIDIA RTX PRO™ 6000 Blackwell GPUs—with the team investing in more GPUs annually to support growing training and AI inference demands. NVIDIA Megatron-Core handles distributed pretraining at scale, delivering throughput and stability for foundation models with hundreds of billions of parameters. For the post-training phase—where models learn to follow instructions and align with e-commerce-specific behavior—the team uses the NVIDIA NeMo framework for instruction tuning and preference alignment. At inference, NVIDIA TensorRT-LLM serves the Compass models in production.

On top of this AI infrastructure, Shopee developed a five-stage training pipeline progressing from pretraining through annealing, instruction following, preference alignment, and agentic reinforcement learning. The pipeline incorporates billion-scale proprietary data mined from product titles, reviews, and ratings, along with novel alignment methods—including OTPO, which dynamically weights tokens by semantic importance, and SPEC-RL, which accelerates on-policy reinforcement learning by 2–3x through speculative rollout reuse. A mixture-of-experts architecture with Multi-Token Prediction drives further throughput gains.

The result: an in-house specialized model that handles language-specific tasks with high accuracy. On Shopee’s EcomEval benchmark, Compass v3.5 exceeds GPT-4o, GPT-4.1, Qwen3-235B, and DeepSeek v3.1 across all five regional ecommerce task categories—scoring 94% on Ecom QA versus 85% for GPT-4o, 92% on Shopping Concept versus 82%, and 84% on User Understanding versus 76%. The reasoning-focused Compassmax v3.5 variant scores 93 on Arena Hard v2 and 83 on AIME25—matching or exceeding proprietary reasoning models like GPT-5-Thinking.

In production, Compass now handles the majority of Shopee’s AI traffic. Monthly API tokens grew from 3 billion to 340 billion in eight months—a 113x increase. The model powers search and recommendations (driving higher conversion and ad revenue), anti-fraud detection (50x efficiency over manual review, 90% lower processing costs), vision-based parcel recovery in the supply chain (saving millions annually), and LLM-driven customer service that replaced rigid scripted chatbots with context-aware, multilingual interactions.