Idea
A large-scale multilingual LLM platform tailored for Southeast Asian e-commerce businesses to enhance localized customer engagement and operations
Research Paper
Core Innovation
This paper presents Compass-v3, a 245B parameter Mixture-of-Experts model optimized for Southeast Asian e-commerce. It innovates by using fewer but larger experts with hardware-efficient optimizations to maximize GPU utilization. The introduction of Optimal-Transport Direct Preference Optimization (OTPO) improves instruction adherence in commerce-specific tasks, outperforming existing large models in benchmarks.
Market Size (TAM)
$2–10B TAM, $1–2B SAM; assumption: growing Southeast Asian e-commerce market with increasing AI adoption for multilingual support.
Potential Customers & Pain Points
- E-Commerce Platforms Needing Multilingual AI Support
- Southeast Asian Retailers Lacking Localized Language Models
- AI Developers Seeking Efficient Large-Scale Models for Commerce
- Marketplaces Struggling with Low-Resource Language Coverage
Business Model
API and platform licensing to e-commerce companies and AI developers with tiered pricing based on usage and language support.
Competitive Landscape
- GPT-4
- Google PaLM
- Anthropic Claude
Implementation Challenges
- High computational resource requirements
- Data privacy and localization challenges
- Adoption resistance in low-resource markets
Validation Strategy
- Deploy pilot with regional e-commerce partners
- Benchmark against GPT-4 on commerce tasks
- Collect user feedback on language and instruction adherence
Research Paper Overview
Compass-v3: Scaling Domain-Specific LLMs for Multilingual E-Commerce in Southeast Asia
Summary
Compass-v3 is a 245B parameter Mixture-of-Experts language model optimized for Southeast Asian e-commerce, trained on 12T tokens of multilingual and synthetic data. It uses fewer but larger experts with hardware-efficient optimizations to maximize GPU utilization. The model introduces Optimal-Transport Direct Preference Optimization (OTPO) to improve instruction adherence in commerce-specific tasks. Compass-v3 outperforms GPT-4 and other large models in e-commerce benchmarks and supports multiple low-resource Southeast Asian languages while maintaining strong general performance. It is deployed at Shopee, handling over 70% of their LLM traffic.