Shopee's Compass-v3 AI Model Masters Multilingual E-Commerce Across Southeast Asia

Researchers at Shopee have unveiled Compass-v3, a large language model engineered specifically for the unique challenges of e-commerce in Southeast Asia. The model addresses a critical limitation of general-purpose AI systems: while they perform well on broad tasks, their effectiveness drops sharply when applied to specialized domains with complex, noisy, and rapidly changing data.

Model Architecture & Scale#

Compass-v3 leverages a Mixture-of-Experts (MoE) architecture with 245 billion total parameters, of which 71 billion activate per token. Rather than spreading computation across many small experts, the team opted for fewer but larger experts paired with hardware-efficient optimizations—including intra-node expert parallelism and a custom memory copy operator—to push GPU utilization to its practical limit. This design choice balances computational complexity with real-world deployment constraints.

Training & Alignment#

The system was trained on 12 trillion tokens drawn from carefully curated multilingual corpora and large-scale synthetic e-commerce instruction data, using a mixed-training strategy to balance breadth and depth. To improve how well the model follows commerce-specific instructions, the researchers introduced Optimal-Transport Direct Preference Optimization (OTPO), a technique that captures fine-grained distinctions at the token level and strengthens adherence to domain-specific requirements.

Multilingual Strength & Real-World Impact#

Compass-v3 demonstrates robust performance across six low-resource Southeast Asian languages—Indonesian, Thai, Filipino, Vietnamese, Malay, and Tagalog—as well as Portuguese, while maintaining competitive results on standard benchmarks. In evaluations, it surpasses DeepSeek-V3.1, GPT-4 variants, and Qwen3-235B on e-commerce-specific tasks. The model has already been deployed at scale within Shopee's platform, where it now accounts for over 70% of total LLM traffic and is progressively replacing OpenAI integrations, demonstrating its practical value in a high-stakes commercial environment.