Shopee Builds Custom LLM with NVIDIA AI Factory: A Turning Point for Southeast Asian AI

Have you ever wondered what is really happening behind every time you search for a product on Shopee, every "tailored" recommendation that appears on the homepage, or every quick response from the support chatbot? The answer lies in something that few people see: artificial intelligence.

But not just any generic AI. This is a stream of AI built specifically for Southeast Asian e-commerce, understanding the language nuances, culture, and user behavior in the region. And recently, Shopee revealed how they built it with the support of NVIDIA AI Factory—a major milestone in the e-commerce giant's AI journey.

Join me, SilverZ, having witnessed technology change the world for 20 years, in analyzing this event from its roots. Because this is not just a story about an AI model. This is a story about how a Southeast Asian company built its own "brain."

The Gap That General-Purpose AI Cannot Fill

Shopee serves hundreds of millions of active monthly users across eight markets in Southeast Asia, Taiwan, and Brazil, with total annual gross merchandise volume exceeding $136 billion. As the platform integrated AI into search, recommendations, fraud detection, and customer service, a structural gap emerged.

General-purpose AI models are great at English tasks, but drop significantly when handling specialized multilingual work—precisely what defines e-commerce in the region. Shopee researchers call this the "jagged frontier"—the same model that scores high on English tests struggles with Indonesian shopping queries, Thai product categorization, and Vietnamese customer intent.

To quantify this gap, Shopee's research team built EcomEval—the first evaluation benchmark dedicated to Southeast Asian e-commerce. The results were alarming: even top models showed a significant decline across multilingual tasks, with some open-source models dropping below 52% in key categories.

For a platform processing billions of queries a month in six languages, that gap translates to lost revenue, prolonged support times, and missed fraud cases. The need for a specialized model became increasingly urgent.

The Birth of Compass: Shopee's Own AI "Brain"

And that is when Compass was born—a specialized, multilingual large language model family built with NVIDIA AI Factory.

Compass is not a general-purpose model "repurposed" from global LLMs. It was built from scratch for the Southeast Asian context, understanding the regional language nuances, culture, and user behavior. According to NVIDIA, Compass outperforms top general-purpose models in e-commerce tasks and has delivered measurable impact.

The numbers are truly staggering:

• Fraud detection is 50 times faster than manual checks.
• Processing costs are reduced by 90% compared to general production.
• Monthly API token volume surged from 3 billion to 340 billion in eight months—a 113-fold increase.

Compass now handles the vast majority of Shopee's AI traffic across search, recommendations, fraud detection, parcel recovery, and customer service. This is no longer a laboratory experiment. This is real production at the scale of hundreds of millions of users.

Compass-v3: The Architecture Behind the Power

To understand why Compass is so effective, we need to look at the technical architecture behind it.

Shopee pre-trained Compass-v3—a domain-specific mixture-of-experts (MoE) model with 245 billion total parameters and 71 billion active parameters per token, specifically designed for Southeast Asian e-commerce.

The model was trained on 12 trillion tokens from a carefully curated multilingual corpus, including Indonesian, Thai, Filipino, Malay, Tagalog, and Portuguese, alongside large-scale synthetic e-commerce instructions using a mixed training strategy.

Building on that foundation, Shopee developed a five-stage training pipeline, progressing from pre-training through annealing, instruction tuning, preference alignment, and agentic reinforcement learning. This pipeline combines billions-of-samples proprietary data mined from product titles, reviews, and ratings, alongside novel alignment methods—including OTPO, which automatically adjusts token weights based on semantic importance, and SPEC-RL, which accelerates on-policy reinforcement learning by 2-3 times through speculative rollout reuse.

The mixture-of-experts architecture with Multi-Token Prediction further drives higher throughput performance. This is a system designed not only to be smart but also fast and cost-effective.

How Compass is Transforming Shopee

Impressive numbers only matter when they translate into real-world value. And Compass is doing just that across Shopee's operational areas.

Search and Recommendations: Boosting Conversion and Ad Revenue

When users search for products in Vietnamese, Indonesian, or Thai, Compass understands the true intent behind the query—not just keywords, but also context, nuance, and latent needs. The result is more accurate product suggestions, a smoother shopping experience, and higher conversion rates. This is good not only for buyers but also for sellers and Shopee itself.

Fraud Detection: 50x Faster, 90% Cost Reduction

In e-commerce, fraud is a persistent pain point. Compass has changed the game by detecting fraud 50 times faster than manual checks, with a 90% reduction in processing costs. This means Shopee can better protect users and sellers while significantly cutting operational expenses.

Parcel Recovery: Saving Millions of Dollars Annually

In the supply chain, Compass uses computer vision for parcel recovery—a complex process requiring high precision. This system is saving Shopee millions of dollars every year while improving the delivery experience for customers.

Customer Service: From Rigid Chatbots to Intelligent Conversation

Instead of rigid script-based chatbots, Compass delivers multilingual, context-aware interactions. Users can ask questions in natural language and receive tailored, prompt answers. This is the difference between a machine and an assistant that truly understands you.

EcomEval: The First Evaluation Benchmark for Southeast Asian E-Commerce

One of Shopee's most important contributions to the regional AI community is EcomEval—the first evaluation benchmark dedicated to Southeast Asian e-commerce.

Before EcomEval, there was no standard to measure the performance of AI models on regional multilingual e-commerce specific tasks. Developers had to rely on English tests that did not reflect actual real-world challenges. EcomEval changes that by providing an accurate and relevant metric for the Southeast Asian context.

This is a vital contribution not just for Shopee but for the entire regional AI ecosystem. When other companies want to build AI for the Southeast Asian market, they now have a standard to reference.

Why This Story Matters to You

You might be thinking: "I'm a Shopee user, why should I care about their AI infrastructure?"

Let me tell you, it is heavily relevant.

First, your shopping experience is improving every day thanks to Compass. When you search for a product and get accurate results, that's Compass. When you get product recommendations matching your preferences, that's Compass. When you are protected from fraudulent transactions, that's Compass too. This AI is quietly making your experience better, faster, and safer.

Second, this is proof that Southeast Asian AI is maturing. For years, companies in the region depended on global AI models trained primarily in English. Compass proves that Southeast Asian companies can build their own AI that understands their own language and culture. This is an important step forward for the region's digital sovereignty.

Third, EcomEval opens opportunities for the entire ecosystem. When there is a standard to measure AI performance for Southeast Asian e-commerce, other companies—from startups to large enterprises—can use it to develop better products. This is a spillover benefit for the entire region.

When Southeast Asia Builds Its Own "Brain"

I have witnessed many tech waves hit Southeast Asia. From broadband internet, to smartphones, to e-commerce, to digital payments. Each wave brought opportunities and challenges. But this AI wave has a distinct difference: it allows companies in the region to build core capabilities themselves instead of just consuming technology from outside.

Compass is a prime example. Shopee didn't just use OpenAI or Google APIs. They built their own model, with their own data, for their own users. That is the difference between a consumer and a creator.

And in a region with over 600 million internet users, six major languages, and thousands of diverse cultures, having an AI that understands that diversity is not just a competitive advantage. It is a prerequisite to truly serve users.

Compass is a bold move. And it raises a question for all other companies in the region: Will you build your own "brain," or keep borrowing?

What do you think about a Southeast Asian company like Shopee building its own AI model instead of relying on global solutions? Is it a strategic step opening a native AI era for the region, or an expensive gamble while general-purpose models become increasingly powerful and cheaper? If you are a Vietnamese business looking to apply AI, what will you learn from Shopee's strategy—and where will you start? Share your views below, because in this AI race, those who dare not bet will forever remain spectators!