# Compass (Shopee language model)

> Source: https://aiwiki.ai/wiki/compass_shopee
> Updated: 2026-08-20
> Fact-checked: 2026-08-20
> Categories: Large Language Models, Mixture of Experts, Natural Language Processing
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Compass (Shopee language model)." aiwiki.ai, 20 Aug 2026. https://aiwiki.ai/wiki/compass_shopee
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Compass is a proprietary family of [large language models](https://aiwiki.ai/wiki/large_language_model) developed by Shopee and its parent company, Sea, for Southeast Asian languages and [e-commerce](https://aiwiki.ai/wiki/e-commerce) tasks. Public descriptions span the dense CompassLLM system reported in 2024, the sparse Compass-v2 and Compass-v3 models reported in 2025, a reasoning branch called CompassMax-V3-Thinking, and a production model that Sea called Compass Max v3.5 in 2026.[1][2][3][4][6] The family is distinct from OpenCompass and from unrelated geospatial, robotics, and navigation projects named Compass or COMPASS.

Compass is principally an internal model family rather than a public model release. Shopee has published research manuscripts about its design, training, and company-run evaluations, but its official Hugging Face organization listed no public models or datasets as of August 2026.[5] The public record also uses `Compass-v3`, `CompassMax-V3`, `CompassMax-V3-Thinking`, `Compass Max v3.5`, `Compass v3.5`, and `Compassmax v3.5` without publishing a complete mapping among those names. Consequently, the papered checkpoints and later serving names cannot all be treated as interchangeable.

## Development history

| Name in source | Public documentation | Reported scale | Documented role |
|---|---|---:|---|
| CompassLLM, CompassLLM-SFT, CompassLLM-DPO | April 2024 arXiv submission; manuscript dated December 2023 | 7B in the public evaluation | Dense multilingual base, instruction-tuned, and preference-tuned models [1] |
| Compass-v2 | April 2025 technical report | 30B total, 5B active | Fine-grained mixture-of-experts model with internal API deployment [2] |
| Compass-v3 | September 2025 technical report | 245B total, 71B active | Larger-expert MoE for multilingual commerce tasks [3] |
| CompassMax-V3-Thinking | December 2025 arXiv submission | Described only as hundred-billion-scale | Long-chain-of-thought reasoning model produced through tuning, model merging, and reinforcement learning [4] |
| Compass Max v3.5 and variants | Sea announcement and NVIDIA case study in 2026 | Sea describes the main model as 245B | Models reported as live across Shopee [6][7] |

The initial manuscript's experiments identify a 7B CompassLLM model, while the later Compass-v2 report benchmarks a 13B system labeled `Compass-v1`. The first generation therefore appears to have included more than one internal size or checkpoint, although the papers do not document the 13B system's training in detail.[1][2] The v2 paper also shows intermediate labels such as `V1.2` and `V1.3` in a data-volume figure without presenting them as separately released models.[2]

## First-generation CompassLLM

The first report focused on English, Chinese, and Indonesian, treating Indonesian as a [low-resource language](https://aiwiki.ai/wiki/low_resource_languages) relative to the English and Chinese material available for model development. Shopee described a 1.7-trillion-token pretraining corpus whose final language shares were 85.3 percent English, 9.9 percent Chinese, and 4.8 percent Indonesian. It combined Common Crawl, C4, Wikipedia, WebText, academic material, books, and code after filtering, deduplication, language identification, contamination mitigation, and resampling.[1]

CompassLLM used a decoder-only Transformer derived from the LLaMA design, with weight tying, RMSNorm, SwiGLU, rotary position embeddings, and an 80,000-entry SentencePiece byte-pair vocabulary assembled for the three target languages. Shopee used [curriculum learning](https://aiwiki.ai/wiki/curriculum_learning) along two axes: sequence length increased during training, and the share of lower-resource-language material rose over time. The report says training took about 24 days and 160,000 compute-hours on 35 nodes containing eight NVIDIA A100 80GB GPUs each, for 280 training GPUs, with two spare nodes. Megatron-DeepSpeed provided the distributed training framework.[1]

Post-training produced CompassLLM-SFT through [supervised fine-tuning](https://aiwiki.ai/wiki/supervised_fine-tuning) and CompassLLM-DPO through [direct preference optimization](https://aiwiki.ai/wiki/direct_preference_optimization_dpo). Although the report discusses learning from human preferences, its documented pair-construction process had GPT-4 score candidate responses. The preference data should therefore not be understood as exclusively direct human labels.[1]

The base model was pretrained with a 2,048-token sequence length. Its reported 128K context support came from inference-time techniques including dynamic NTK interpolation, LogN attention scaling, and StreamingLLM, rather than from pretraining at 128K. Evidence in the paper consisted of PG-19 perplexity curves and line-retrieval experiments. This was narrower than the multi-task long-context evaluations normally needed to establish reliable 128K behavior.[1]

## Compass-v2

Compass-v2 changed the family from a dense model to a [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) architecture trained from scratch. It had 30 billion total parameters and five billion active parameters. Forty-eight specialized experts were paired with two shared experts; for each token, the router selected four specialized experts in addition to the shared experts. The model used grouped-query attention, RMSNorm, SwiGLU, and rotary position embeddings. Its initial 4,096-token context was extended to 32,768 through further training.[2]

Pretraining comprised eight trillion tokens of broader multilingual data followed by four trillion tokens of higher-quality annealing data. The report also describes a smaller long-context training stage but does not fold a clear additional token count into the 12-trillion total. Shopee built a 180,000-entry tokenizer for English, Chinese, regional languages such as Indonesian, Thai, Vietnamese, Malay, and Tagalog, Portuguese, and a group of European languages. E-commerce material combined external web mining with internal product descriptions, reviews, and other platform text.[2]

Post-training used six million instruction samples in two fine-tuning stages, followed by preference optimization. Shopee described Compass-v2 as a hybrid reasoning model able to provide either a short answer or a longer reasoning response without routing users to two separate models. It was deployed through CAP, Shopee's internal model-as-a-service platform, which offered a web portal and an API. The report explicitly characterized the Compass API as an internal company release. Listed applications included search, recommendation, video chat, and livestream support.[2]

## Compass-v3 and the reasoning branch

Compass-v3 increased the scale to 245 billion total parameters and 71 billion active in each forward pass. Its MoE layers contained 16 comparatively large experts, with four selected for each token. The report argues that this fewer-but-larger design better matched Shopee's hardware profile than architectures built from many small experts. It also added multi-token prediction, intra-node expert parallelism, a custom memory-copy optimization, and expert-aware FP8 quantization.[3]

Shopee reported a 12-trillion-token pretraining corpus. To choose its composition, the team divided candidate data into 16 language and domain shards, sampled 512 mixtures, trained two-million-parameter proxy models for one billion tokens per mixture, and checked the selected mixture with a 1.3-billion-parameter model trained on 150 billion tokens. Training then shifted toward commerce and regional-language material, reasoning data, and long-context data, with context extended first to 32K and then to 128K.[3]

Instruction tuning consolidated v2's two stages into one. The report gives an approximate size of 2.81 million multilingual instruction instances for Indonesian, Malay, Thai, Vietnamese, Tagalog, and Portuguese in addition to English and Chinese. It describes data drawn from open corpora, translation, multi-agent synthesis, business requirements, internal serving logs, and deployment feedback. For alignment, Shopee introduced Optimal Transport Preference Optimization, a token-weighted form of DPO intended to emphasize differences between preferred and rejected responses. The team also used a commerce-specific reward model.[3]

CompassMax-V3-Thinking extended the line toward deliberate reasoning. Its report describes a cold-start stage using long chain-of-thought data, merging with CompassMax-V3 instruction-tuned checkpoints, and two phases of [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning). The first phase emphasized mathematics, code, and instruction following; the second added commerce, tool use, and general question answering. Named techniques included multi-stage zero-variance elimination, entropy importance sampling policy optimization, Router Replay to reduce training-serving routing differences, generative reward modeling, FP8 rollouts, length-aware scheduling, and overlap between token generation and reward computation.[4]

The Thinking report calls its model a hundred-billion-scale MoE but does not restate exact total and active parameter counts. Sea later called Compass Max v3.5 a 245-billion-parameter model, while NVIDIA's case study called the general serving model `Compass v3.5` and a reasoning model `Compassmax v3.5`.[6][7] Those sources do not establish that CompassMax-V3-Thinking and the v3.5 reasoning service are an identical checkpoint.

## Evaluation record

Every public Compass result was produced or commissioned by Shopee, and no downloadable checkpoint was available for independent reproduction. The first report compared its 7B models with other openly described 7B systems on academic tasks. It also used 165 new English and Indonesian questions evaluated by GPT-4, including position-swapped comparisons, and 161 unsafe prompts per language assessed with an automated judge. The authors acknowledged a low safety score for the base scale and the presence of unsafe material in training data.[1]

Compass-v2 evaluation combined public datasets with private multilingual, commerce, and business sets. Standard tasks used `lm-evaluation-harness`; open-ended internal responses were judged by GPT-4, with manual review for selected problematic cases. Models did not share one decoding configuration: Compass, Qwen, LLaMA, and regional baselines had different temperature and sampling settings, while proprietary APIs used their defaults. Shopee's reported advantage therefore describes its own evaluation setup, not a controlled third-party comparison.[2]

Compass-v3 followed a similar pattern. Public tasks were supplemented by private multilingual and commerce questions, and GPT-4.1 judged open-ended items. The report used different generation settings for Compass, Qwen, and commercial APIs. It reported that Compass-v3 led its private commerce average, but the public commerce table contains an internal labeling problem: the prose assigns 49.50 to Shopping MMLU and 85.14 to ECInstruct, while the table layout associates the numbers with the opposite rows. Neither score can be repeated uncritically without resolving that discrepancy.[3]

A later NVIDIA case study presents a Shopee-created benchmark called EcomEval and says Compass v3.5 led GPT-4o, GPT-4.1, Qwen3-235B, and DeepSeek-V3.1 across five commerce categories. It also reports Arena Hard v2 and AIME25 results for the reasoning variant.[7] The charts are marked as courtesy of Shopee, so these remain company results shown by an infrastructure partner rather than independent [benchmark](https://aiwiki.ai/wiki/benchmark) verification.

## Production deployment

The Compass-v3 paper said the family had begun replacing [OpenAI](https://aiwiki.ai/wiki/openai) traffic and accounted for more than 70 percent of Shopee's total LLM use at the time of writing.[3] A 2026 [NVIDIA](https://aiwiki.ai/wiki/nvidia) customer case study later said Compass handled a majority of Shopee's AI traffic. It reported that monthly Compass API volume rose from three billion tokens in April 2025 to 340 billion in November, an increase of about 113 times. The same case study attributed search and recommendation, customer service, anti-fraud review, and parcel-recovery applications to the family.[7]

NVIDIA also reported anti-fraud review that was 50 times faster than manual processing and cost 90 percent less, plus annual savings of millions from parcel recovery.[7] These figures were supplied in a customer success story and have not been independently audited. They should not be generalized into performance or cost guarantees for other workloads.

The partner case study says Shopee's model infrastructure included A100, H100, and RTX PRO 6000 Blackwell GPUs, Megatron-Core for pretraining, NeMo for post-training, and [TensorRT-LLM](https://aiwiki.ai/wiki/tensorrt_llm) for serving.[7] It describes the broader production estate, not a claim that every listed GPU type trained every Compass generation.

In April 2026, Sea announced an AI Centre of Excellence in Singapore and identified Compass Max v3.5 as a 245-billion-parameter foundation model already used across Shopee. Sea said the variants were more responsive and cost a fraction of commercial services.[6] The Business Times independently reported the model's integration while attributing that comparative cost statement to Sea rather than presenting it as a tested finding.[8]

## Availability and reproducibility

Shopee's official Hugging Face organization labels CompassLLM as a self-developed text-model series, but it displayed zero public models and zero public datasets as of 19 August 2026.[5] The technical reports do not identify an official download repository for model weights, tokenizers, training code, private evaluation sets, or serving code, and they do not provide a model-weight license.[1][2][3][4] CAP was described as an internal access system, not a public API.[2]

These omissions limit reproducibility. The first-generation and v2 manuscripts also contain an unusually broad Shopee disclaimer stating that the company does not warrant the accuracy or sufficiency of the presented information, conclusions, research design, or methodology.[1][2] The v3 and Thinking reports provide considerably more systems detail, but their central training corpora, request logs, human annotations, preference pairs, internal benchmarks, and deployment telemetry remain private.[3][4]

## Significance and limitations

Compass is one of a set of models built to address the uneven representation of Southeast Asian languages in general-purpose systems. ABI Research included the first generation alongside SEA-LION, SeaLLM, and Sailor in its industry survey, identifying CompassLLM-SFT and CompassLLM-DPO as 7B systems trained on 1.7 trillion tokens for English, Chinese, and Indonesian.[9] A broader Carnegie Endowment study documented rapid growth in Southeast Asian language-model initiatives and emphasized that openness depends separately on access to weights, code, data, licenses, documentation, and APIs.[10] Under that framework, Compass is publicly documented but operationally closed.

The family's strongest evidence concerns its existence, technical lineage, and internal deployment. Its weakest evidence concerns comparative quality and business impact. Private test sets prevent outside inspection of sampling, contamination, judge reliability, and representativeness. Different decoding settings complicate model comparisons. The v3 commerce-table inconsistency introduces an additional reporting problem. Finally, the changing public names make it unsafe to transfer a result from one Compass checkpoint or serving alias to another unless the source explicitly identifies the version.

## References

1. Maria, Sophia, and Shopee Large Language Model Team. Compass: Large Multilingual Language Model for South-east Asia. arXiv:2404.09220, 2024. https://arxiv.org/abs/2404.09220
2. Maria, Sophia, and Shopee LLM Team. Compass-v2 Technical Report. arXiv:2504.15527, 2025. https://arxiv.org/abs/2504.15527
3. Maria, Sophia, and Shopee LLM Team. Compass-v3: Scaling Domain-Specific LLMs for Multilingual E-Commerce in Southeast Asia. arXiv:2509.09121, 2025. https://arxiv.org/abs/2509.09121
4. Zeng, Anxiang, et al. Each Prompt Matters: Scaling Reinforcement Learning Without Wasting Rollouts on Hundred-Billion-Scale MoE. arXiv:2512.07710, 2025. https://arxiv.org/abs/2512.07710
5. ShopeeLLM. Shopee official Hugging Face organization. Accessed 19 August 2026. https://huggingface.co/ShopeeLLM
6. Sea Limited. Sea Establishes AI Centre of Excellence in Singapore to Drive AI-Native Innovation and Capability Building. 20 April 2026. https://www.sea.com/news/394
7. NVIDIA. Shopee's LLMs Deliver Frontier-Level Performance With the NVIDIA AI Factory. 2026. https://www.nvidia.com/en-us/case-studies/shopee/
8. Fok, Jermaine. Sea to launch new AI centre to drive innovation and capability-building in Singapore. The Business Times, 20 April 2026. https://www.businesstimes.com.sg/companies-markets/sea-launch-new-ai-centre-drive-innovation-and-capability-building-singapore
9. ABI Research. Multilingual/Bilingual Large Language Models: Tailoring AI Applications for Southeast Asia. 2024. https://www.abiresearch.com/blog/multilingual-bilingual-large-language-models-llms
10. Sastrosubroto, Ashwin Sasongko, and Kat Duffy. Speaking in Code: Contextualizing Large Language Models in Southeast Asia. Carnegie Endowment for International Peace, 6 January 2025. https://carnegieendowment.org/research/2025/01/speaking-in-code-contextualizing-large-language-models-in-southeast-asia