Shopee Brings Most of Its AI In-House With Compass, Its Own E-commerce LLM Built With NVIDIA, as Monthly Tokens Grow 113x in Eight Months
Shopee has built Compass, a multilingual e-commerce LLM, with NVIDIA. We explain why general-purpose LLMs falter in Southeast Asian e-commerce, what the EcomEval benchmark measures, how Compass is used in search and customer service, and the caveats behind the numbers.
Key Takeaways
- Shopee has built Compass, a multilingual LLM for Southeast Asian e-commerce, in house on NVIDIA's computing platform. It already handles the majority of Shopee's AI traffic across search, recommendations, fraud detection and customer service, and its monthly token volume grew 113x in eight months, from 3 billion to 340 billion
- The driver is the "jagged frontier": general-purpose LLMs that score well in English lose accuracy on Indonesian shopping queries and Thai product categorization. Shopee first built EcomEval, a dedicated benchmark drawn from real customer queries and transaction logs, to measure that gap before committing to its own model
- Most of the published numbers are self-reported by Shopee and NVIDIA, and the comparisons are mostly against previous-generation models. The lesson for merchants is still clear: before handing work to agents, build a way to measure where general-purpose models fail in your own language and product categories
Shopee Builds Its Own LLM, Compass, With NVIDIA
Southeast Asian e-commerce giant Shopee has built Compass, a family of specialised, multilingual Large Language Models (LLMs), with the NVIDIA AI Factory.
www.frontier-enterprise.comShopee, Southeast Asia's largest e-commerce platform, now runs most of its internal AI workload on Compass, a multilingual LLM it pretrained itself specifically for e-commerce. According to a case study disclosed by NVIDIA, Compass was developed and runs on an "AI Factory" built from NVIDIA GPUs and software.
The growth has been steep. The monthly API token volume processed by Compass rose from 3 billion to 340 billion in eight months, a 113x increase. Tokens are the units of text a model reads and writes, and serve as a rough measure of workload. Frontier Enterprise reports that its uses now span search and recommendations, fraud detection, parcel recovery in logistics and customer service.
The timing coincided with NVIDIA AI Day, held in Singapore on September 23. There, parent company Sea Limited said it would be the first enterprise in Southeast Asia to adopt NVIDIA's next-generation Vera Rubin platform. The size of the deployment, its timing and the financial terms were not disclosed.
Shopee's scale is worth keeping in mind. It serves hundreds of millions of monthly users across eight markets in Southeast Asia, Taiwan and Brazil, with annual gross merchandise value (GMV) above $136 billion. Li Kuangli, senior director of partnerships and strategy at Sea, said Shopee has more than 400 million annual active buyers placing over 15 billion orders a year.
Why General-Purpose LLMs Stumble on Indonesian Shopping
Shopee's researchers call the uneven capability of general-purpose LLMs the "jagged frontier". A model that scores well on English benchmarks suddenly loses accuracy on Indonesian shopping queries, Thai product categorization and Vietnamese customer intent. According to NVIDIA, some open source models scored below 52% on key categories.
At a scale of billions of queries a month across six languages, that gap turns directly into losses. Misreading intent means lost sales in search, slower resolution in support and missed fraud.
What Shopee did first was not build a model but build a way to measure. EcomEval, the benchmark it published as a paper in October 2025, covers 37 tasks in six categories, eight of them multimodal tasks involving images. Most questions are drawn from real customer queries and transaction logs, and the benchmark covers seven languages: English and Chinese plus Indonesian, Vietnamese, Thai, Malay and Portuguese. Reference answers were drafted by large models and then reviewed and revised by more than 50 expert annotators with e-commerce and multilingual expertise, in a semi-automated pipeline.
Existing e-commerce benchmarks focused on English and Chinese, often relied on synthetic or curated data, and left out practical tasks such as pre-purchase product guidance and post-purchase after-sales service.
The paper's results contain several points merchants should read closely.
- Every model performs worse on tasks in low-resource languages such as Indonesian and Malay than in English. The gap between models also varies by language: the difference between first and second place is 0.8 points in English but widens to 3.37 points in Indonesian
- On query-product matching and search relevance, models tend to overlook product categories and the intended buyer
- On pre-purchase guidance and after-sales questions, scores stay between 62% and 78%
The second and third points carry real weight for agentic commerce. For an AI agent to shop on someone's behalf, it first has to read correctly what that person is looking for. If general-purpose models fall short exactly there, no amount of work on payments or checkout will make transactions flow correctly.
A 245-Billion-Parameter Vertical MoE and a Five-Stage Training Pipeline
The latest generation, Compass-v3, is a vertical-domain mixture-of-experts (MoE) model designed for Southeast Asian e-commerce. An MoE model contains many specialized sub-networks and activates only some of them for each input. According to the technical paper, it has 245 billion total parameters with 71 billion active per token, and it uses fewer but larger experts to raise GPU utilization.
The model was trained on 12 trillion tokens, mixing multilingual corpora including Indonesian, Thai, Filipino, Malay, Tagalog and Portuguese with large volumes of synthetic e-commerce instructions. It also drew on billion-scale proprietary data mined from product titles, reviews and ratings.
As NVIDIA describes it, training proceeds in five stages: pretraining, annealing (a finishing phase that raises the weight of e-commerce and multilingual data), instruction following, preference alignment and agentic reinforcement learning. The final stage has the model improve through trial and error toward a reward on multi-step tasks, which bears on the quality of its behavior in conversational selling and support.
Two in-house methods are also highlighted. OTPO performs preference learning while weighting each token by its semantic importance, and SPEC-RL speeds up reinforcement learning by 2 to 3x by reusing earlier rollouts. SPEC-RL was published in a paper co-authored with Xiamen University and others, with its code released as well.
What Compass Handles in Production
For practitioners, the more pressing question is where in the transaction flow Compass is used. Here is what NVIDIA and Sea have disclosed.
| Use case | What Compass does | Reported impact |
|---|---|---|
| Search and recommendations | Reads multilingual shopping intent and matches and ranks products | Higher conversion and ad revenue (size of lift not disclosed) |
| Fraud detection | Flags fraudulent transactions and behavior | 50x the efficiency of manual review, 90% lower processing costs |
| Parcel recovery | Vision-based parcel recovery in the supply chain | Millions of dollars saved annually (breakdown not disclosed) |
| Customer service | Replaces scripted chatbots with context-aware, multilingual responses | AI agents handle over 80% of customer queries (per Sea) |
| Overall | Handles the majority of Shopee's AI traffic | Monthly API tokens up from 3 billion to 340 billion in eight months (113x) |
What stands out is that Compass has spread as a replacement for external models. The Compass-v3 paper states that in Shopee's production environment it is gradually replacing OpenAI's traffic and now accounts for more than 70% of total LLM usage. The 113x growth in monthly tokens is best read as including workloads moved from external APIs to the in-house model, not just new use cases.
The same shift shows up in earnings commentary. On the Q2 2026 earnings call in August, management said it had been rolling out a generative algorithm for search and recommendations and had seen a meaningful improvement in conversion. The ad take rate improved by more than 90 basis points year over year, with AI helping match buyer intent to products.
Customer service is being replaced too. According to Li Kuangli, AI agents on Shopee handle over 80% of customer queries without hurting customer satisfaction. NVIDIA says Compass replaced rigid scripted chatbots with context-aware, multilingual interactions.
How to Read the Numbers
Most of the published figures come from Shopee itself or from NVIDIA, which supplies the GPUs. They should be treated as numbers that no third party has verified.
The performance comparisons deserve particular care. The paper reports that Compass-v3 scored 86.56% overall on Shopee's in-house e-commerce evaluation set, ahead of GPT-4.1 at 82.03%, GPT-4o at 76.67% and DeepSeek-V3.1 at 76.55%. But the evaluation data is in-house, and the comparison models are mostly a generation behind as of publication. A company's own model winning a test built around its own operations is, in a sense, to be expected.
The "50x" fraud detection figure is also measured against manual review, not against a general-purpose LLM. The parcel recovery savings, described as millions of dollars a year, come without a specific amount or methodology. The cost of developing and running the model has not been disclosed either, so outsiders cannot judge whether building in house turned out cheaper than using external APIs.
Does This Contradict the OpenAI and Google Partnerships?
Sea signed partnerships with Google in February 2026 and with OpenAI in June. The OpenAI deal made the Shopee app available inside ChatGPT. Partnering with outside AI companies while growing an in-house LLM can look inconsistent at first glance.
The two play different roles, though. The ChatGPT and Google partnerships are about where shoppers come in, the outer entry points. Compass is about the inner brain: how Shopee understands incoming queries, which products it returns and how it responds. The entry points are opened to the large outside players, while the parts tied directly to Shopee's strengths, understanding and matching intent, stay in house.
Sea does not claim to do everything itself. Li said it uses third-party models where they suit the task, adapts open models for its needs and chooses models on pragmatic grounds. Internal applications reach models through a shared gateway that lets them pick whichever fits the job best.
A similar pattern appears at large US retailers. In October 2024 Walmart announced Wallaby, a suite of retail-specific LLMs trained on decades of its own data. Its approach of swapping external and proprietary models on an internal platform closely resembles Sea's. Amazon likewise says its shopping assistant Rufus runs on a custom LLM specialized for shopping, trained on its product catalog, reviews and community Q&A. The larger the platform, the clearer the tendency to own the model responsible for understanding purchase intent.
What Merchants Can Take Away
Only a handful of companies can pretrain a 245-billion-parameter model. It takes the scale of $136 billion in annual GMV and, across Sea, the capacity to run more than 25,000 GPUs in house. This is not something to copy directly.
There are still lessons to take home. The most broadly applicable is the order of operations: first measure where general-purpose models fail in your own language and product categories. Before building its own model, Shopee built an evaluation set from real queries and transaction logs. For a Japanese merchant, the equivalent would be a small evaluation set of Japanese-specific phrasing, model numbers and standards, size notation, and return and after-sales questions, run against candidate models. Product guidance and after-sales service, where scores stayed low even on EcomEval, are the first areas to check.
The second lesson is that product data and reviews are assets for both training and evaluation. Compass's strength rests on proprietary data mined from product titles and reviews. Even for merchants that will never train a model, product data with consistent attributes and well-organized inquiry histories are the foundation for evaluating and tuning external models and for getting agents to read products correctly.
The third concerns model selection. As Sea's shared gateway and Walmart's internal platform show, evaluating and assigning models by use case is more realistic than handing everything to a single general-purpose model.
Conclusion
Compass is a case of a platform concluding that general-purpose LLMs cannot fully handle "understanding purchase intent" in multilingual e-commerce. That understanding is the prerequisite for agentic commerce, and Shopee filled the gap in house. While opening its front door to ChatGPT and Google, Shopee is building a structure in which its own model handles most of the AI processing at the core of the transaction: search, recommendations and customer service.
The next things to watch are third-party verification of performance and how far agentic capabilities expand after Vera Rubin arrives. The paper also lists incorporating images and video and continual training to keep up with changing products and user behavior as next steps. Rely entirely on general-purpose models, or hold your own evaluation criteria? Whatever their size, merchants now need an answer to that question.
