Large Language Models (LLMs) are a type of artificial intelligence model designed to understand and generate human language. Recent developments in LLMs have showcased remarkable capabilities in understanding and generating human language, with applications spanning translation, summarization, coding assistance, question answering, and more.
Many existing LLMs, however, are trained upon massive internet-based datasets, which often has disproportionately large influences from western, industrialized, rich, educated, and democratic (WIRED) societies, as people from non-WIRED societies are less likely to be literate, to use the Internet, and to have their output easily accessed. Such an imbalance in the training data can lead to model outputs that display strong bias in terms of cultural values, political beliefs and social attitudes.
LLMs trained on predominantly WIRED-centric content risk neglecting the linguistic and cultural diversity inherent in non-WIRED populations. These biases become evident not only in examples like cultural references or local idioms, but also in more critical domains such as decision-making, social attitudes and moral reasoning, and which can vary significantly across global communities. By overlooking these variations, mainstream LLMs may inadvertently perpetuate inaccurate assumptions or exclude large segments of the global population.
Our work in SEA-LION, now part of Singapore’s National Multi-Modal Large Language Model project, aims to address these disparities by creating LLMs that cater to under-represented population groups and low resource languages in the SEA region.
SEA-LION is trained on more content produced in Southeast Asian languages like Thai, Vietnamese and Bahasa Indonesia to ensure better representation in data and alignment compared to Western or Chinese models. SEA-LION models understand nuances in SEA languages and demonstrate greater awareness of cultural context specific to the region.
This lowers the bar for governments, industries, and academia that seek LLM solutions that fit local languages and reflect local cultural norms, since WIRED-centric models can pose langauge barriers and misalign with local sensibilities in the SEA region.
Built for Southeast Asia, by Southeast Asia
SEA-LION (Southeast Asian Languages in One Network) is a family of open-source Large Language Models (LLMs) that better understands Southeast Asia’s (SEA) diverse contexts, languages, and cultures.
It is an open-source project anchored by the Products Pillar of AI Singapore. Our work in SEA-LION aims to create LLMs that cater to under-represented population groups and low resource languages in the SEA region. You can read more about our motivations for SEA-LION here.
This site provides information and resources on SEA-LION, including how to access the models, hosting, and how-to guides.
¹ Supervised Fine-Tuning
² Continued Pre-Training
SEA-LION has seen:
In v1, ability to outperform most models based on SEA-HELM (Southeast Asian Holistic Evaluation of Language Models) when it was released
In v2, outperformance for SEA tasks, while retaining credible performance on standard (English) benchmarks
In v2.1, key improvements in conversational abilities across SEA languages, while providing more helpful and contextually appropriate responses to user prompts
In v3, outperforms similar sized open source models, and even some larger models in both general and SEA capabilities
SEA-LION-Embedding: The Vector Foundation
The SEA-LION-Embedding suite provides the semantic foundation for the ecosystem. Released in March 2026, these models represent a significant leap in regional retrieval performance. Tested on the SEA-BED (Southeast Asia Embedding Benchmark), which utilizes human-curated native data rather than machine translations, our embeddings consistently set new state-of-the-art records for 10 regional languages across retrieval, reranking, and semantic textual similarity tasks.
SEA-LION-Guard: The Protector
In v1, building on the sophisticated reasoning and multimodal foundation laid by SEA-LION v4, we are proud to introduce the first generation of SEA-Guard. Released on 4 Feb 2026, SEA-Guard is the dedicated safety counterpart to the SEA-LION family. As our foundational models gained the power to "see" (Multimodality in v4) and "think" (Reasoning in v3.5), the need for a culturally attuned safety layer became paramount.
In v2, we elevate safety from passive moderation to active governance. Built on the Gemma-4 foundation architecture, Gemma-SEA-LION-Guard-v2 (31B, E4B, and E2B) transitions from traditional text filtering to real-time Action Governance, Bring Your Own Policy (BYOP) enforcement, and Multi-Turn Context Evaluation across a massive 256K context window. Engineered specifically for agentic deployments, v2 replaces freeform text responses with calibrated, type-safe decision outputs (content_moderation, rule_applicability, and rule_violation)—enabling software to intercept unauthorized tool calls, enforce custom enterprise rules, and detect multi-turn scam patterns across 8 Southeast Asian languages.
We use a holistic approach to evaluation, including not just traditional Natural Language Processing (NLP) benchmarking tasks (such as sentiment analysis and question answering) but also .
Visit our for more detailed breakdown on:
How SEA-LION compares to other available models along different metrics
What SEA-HELM is and the four key capabilities it is evaluated on: English performance, Proficiency in SEA chat, Instruction-following and Linguistic tasks
What each of these globally recognized metrics mean under SEA-HELM
Transparent and Open Source
We have benefited greatly from the open-source community and believe that efforts to better represent our region will similarly be well served by open-source efforts.
All SEA-LION releases will therefore embrace an open-source ethos under the MIT license as much as possible; however, the exact licensing terms may vary depending on the underlying base model’s restrictions or requirements. For instance, if the model leverages Meta’s Llama3 codebase, it may be bound by the , which places certain restrictions on commercial use. Similarly, the Gemma-based variants may carry different terms. Users should always refer to the Hugging Face model card of each specific SEA-LION model for the most accurate, up-to-date license information.
SEA-LION will also be open and transparent in the following areas throughout this guide:
Pre-Training data
Model training code
Fine-Tuning data
Evaluation benchmarks
We welcome contributions to SEA-LION! Check out the to get started.
Some ways to contribute:
Report bugs and issues
Enhance the documentation
Add more model evaluation tasks and metrics
Train versions of the model in more SEA languages
Check out our also, for possible ways to further enhance and expand the capabilities of SEA-LION together.
If you use SEA-LION in your work, please cite it as:
If you are using SEA-LION v3 for your work, please cite it as:
If you are using SEA-LION embedding for your work, please cite it as:
AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore. Any opinion, finding, conclusion or recommendation expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore, or the National University of Singapore.
We also grateful for the support of the Infocomm Media Development Authority (IMDA) of Singapore.
SEA-LION would not be possible without a growing list of Singapore, regional, and international collaborators. Please see our website for more details.
If you have questions, comments, or issues, please open a GitHub issue or contact us via .
In v3.5, ability to handle reasoning tasks, with the versatility of handling general tasks as well while maintaining similar performance with state-of-the-art models.
In v4, our first multimodal SEA-LION models, extending capabilities beyond text to handle image + text inputs with massive 256K native context windows and specialized regional OCR, while continuing our focus on Southeast Asian languages, culture, and use cases.
In v4.5, rapid specialization of state-of-the-art open foundation models via knowledge distillation and model merging, delivering high-capacity reasoning and agentic tool-use capabilities alongside speed-optimized booster configurations for low-latency production deployment.
In v4.8, extended language coverage across 13 regional languages—including Burmese, Malay, and key sub-national languages—paired with advanced post-training for enhanced multi-turn chat and instruction-following, rigorously benchmarked under the upgraded SEA-HELM suite with deeper cultural competence (KALAHI) and country-entity knowledge evaluation.
31B
256K
SFT¹ of gemma-4-31B-it (Latest)
Instruct
E4B, E2B
128K
SFT¹ of gemma-4-E4B-it, gemma-4-E2B-it (Latest)
Instruct
30B-A3B
262K
CPT²,SFT¹ of NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16
Base, Instruct, GGUF, NVFP4, FP8
120B-A12B
262K
CPT²,SFT¹ of NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16 (Latest)
@misc{sea_lion_2024,
title={SEA-LION (Southeast Asian Languages In One Network): A Family of Large Language Models for Southeast Asia},
author={AI Singapore},
year={2024},
howpublished={\url{https://github.com/aisingapore/sealion}}
}@misc{2504.05747,
title={SEA-LION: Southeast Asian Languages in One Network},
author={Raymond Ng and Thanh Ngan Nguyen and Yuli Huang and Ngee Chia Tai and Wai Yi Leong and Wei Qi Leong and Xianbin Yong and Jian Gang Ngui and Yosephine Susanto and Nicholas Cheng and Hamsawardhini Rengarajan and Peerat Limkonchotiwat and Adithya Venkatadri Hulagadri and Kok Wai Teng and Yeo Yeow Tong and Bryan Siow and Wei Yi Teo and Wayne Lau and Choon Meng Tan and Brandon Ong and Zhi Hao Ong and Jann Railey Montalan and Adwin Chan and Sajeban Antonyrex and Ren Lee and Esther Choa and David Ong Tat-Wee and Bing Jie Darius Liu and William Chandra Tjhi and Erik Cambria and Leslie Teo},
year={2025},
eprint={2504.05747},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2504.05747},
}@misc{2606.03027,
title={SEA-LION-Embedding: Open and Reproducible Text Embeddings for Southeast Asia},
author={Peerat Limkonchotiwat and Raymond Ng and Sarana Nutanong and Jian Gang Ngui},
year={2026},
eprint={2606.03027},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2606.03027},
}@misc{2602.01618,
title={SEA-Guard: Culturally Grounded Multilingual Safeguard for Southeast Asia},
author={Panuthep Tasawong and Jian Gang Ngui and Alham Fikri Aji and Trevor Cohn and Peerat Limkonchotiwat},
year={2026},
eprint={2602.01618},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2602.01618},
}Base, Instruct, GGUF, NVFP4, FP8
E2B
128K
SFT¹ of gemma-4-E2B-it
Instruct, GGUF
27B
262K
SFT¹ of Qwen3.6-27B
Instruct, SpecDecoder, GGUF
4B
128K
SFT¹ of Gemma 3 4B IT
VLM
8B
65K
SFT¹ of Apertus-8B-Instruct-2509
Instruct, Reasoning
4B, 8B
256K
SFT¹ of Qwen3-VL-4B-Instruct, Qwen3-VL-8B-Instruct
VLM
27B
128K
CPT² of Gemma 3 27B IT
Base, Instruct, GGUF, NVFP4, FP8_Dynamic
27B
128K
SFT¹ of Gemma-SEA-LION-v4-27B-IT
VLM
32B
128K
SFT¹ of Qwen3-32B
Instruct, GPTQ-4BIT, GPTQ-8BIT
8B
128K
SFT¹ of Llama-SEA-LION-v3-8B-IT
Reasoning, GGUF
70B
128K
SFT¹ of Llama-SEA-LION-v3-70B-IT
Reasoning, GGUF
9B
8192
CPT² of Gemma2
Base, Instruct, GGUF
8B
128K
CPT² of Llama 3.1 8B
Base, Instruct, GGUF
70B
128K
CPT² of Llama 3.1 70B
Base, Instruct, GGUF
8B
8192
CPT² of Llama3
Base, Instruct, GGUF
3B
2048
Pre-training from scratch
Base
7B
2048
Pre-training from scratch
Instruct
300M, 600M
8K
ModernBERT from scratch
Base-Embedding, Checkpoints
300M, 600M
8K
SFT¹ of SEA-LION-ModernBERT-Embedding
Tuned-Embedding
600M
512
SFT¹ of E5-Large
Tuned-Embedding
4B, 8B
128K
SFT¹ of Qwen-SEA-LION-v4-4B-VL, Qwen-SEA-LION-v4-8B-VL
VLM
8B
128K
SFT¹ of Llama-SEA-LION-v3-8B-IT
Instruct
12B
128K
SFT¹ of Gemma 3 12B IT
VLM
We, the community of contributors and users of SEA-LION, pledge to create a welcoming and inclusive environment for everyone. We are committed to fostering a respectful and harassment-free space where diverse ideas and perspectives can thrive.
To contribute to a positive and inclusive atmosphere, we expect all participants, including contributors, users, and maintainers, to:
Be respectful and considerate: Treat others with kindness, respect, and empathy. Recognize and embrace diversity in backgrounds, experiences, and opinions.
Be inclusive: Welcome and support people of all backgrounds, identities, and abilities. Avoid any form of discrimination or exclusionary behavior.
Listen actively: Pay attention to others' ideas, experiences, and feedback. Be open to constructive criticism and different points of view.
Show empathy: Understand that people may have different cultural norms, communication styles, and perspectives. Be patient and considerate when engaging with others.
Resolve conflicts constructively: Disagreements and conflicts are natural, but we encourage participants to address them in a respectful and solution-oriented manner. Avoid personal attacks and name-calling.
Use clear and inclusive language: Use language that is respectful, inclusive, and considerate of all participants. Avoid offensive, derogatory, or discriminatory language.
The following behaviors are considered unacceptable and will not be tolerated within the SEA-LION community:
Harassment: Any form of harassment, including but not limited to offensive comments, slurs, intimidation, or unwelcome advances, is strictly prohibited.
Discrimination: Discriminatory actions or comments based on race, ethnicity, nationality, gender, gender identity, sexual orientation, disability, religion, age, or any other characteristic will not be tolerated.
Hate speech: Hate speech, promoting violence, or advocating harm towards individuals or groups based on their identity is not allowed.
Personal attacks: Engaging in personal attacks, insults, or trolling of others within the community is unacceptable.
If you witness or experience behavior that violates this code of conduct, please report it promptly to the project maintainers by contacting
All reports will be treated with confidentiality, and the project maintainers will take appropriate action as necessary to address violations. We are committed to providing a safe and welcoming environment for all participants.
Enforcement of this code of conduct will be carried out in a fair and just manner. Depending on the severity and frequency of violations, consequences may include warnings, temporary or permanent bans from the community, or other appropriate actions.
This code of conduct is adapted from the , version 2.0, available at https://www.contributor-covenant.org/version/2/0/code_of_conduct.html.
Common questions we receive at the SEA-LION inbox, grouped by topic. If your question is not covered here, write to us at sealion@aisingapore.org or use the Collaborate form on sea-lion.ai.
SEA-LION (Southeast Asian Languages In One Network) is a family of open-source large language models built for Southeast Asia's languages and cultural contexts. The models are developed by AI Singapore.
You have a few options:
Test the models directly in our Playground.
Chat with SEA-LION on our Telegram bot.
Download the models from .
Inference the models through major cloud platforms (see "Running SEA-LION in production" below).
Yes. SEA-LION is open-source and free to use, including for commercial purposes. You can download and self-host the models from Hugging Face, or access them through cloud platform partners.
You can generate an API key through the SEA-LION Playground. Documentation is available at .
Our free APIs are intended for proof-of-concept (POC) development and are rate-limited (10 requests per minute). They are not designed for production workloads.
We are not able to raise the limits on the free POC tier. If you need higher throughput or production-grade reliability, inference SEA-LION through one of our cloud platform partners, where you can scale usage according to your needs.
The free API is rate-limited and meant for POC testing, so heavy or sustained use may hit limits. For continued or production use, move to a cloud platform partner. If you believe you are seeing an actual outage rather than a rate limit, write to us with the details.
For production, inference SEA-LION through major cloud platforms rather than the free POC API. The models are available on hyperscalers including AWS (Bedrock) and Google Cloud (Vertex AI), among others. See our for the current list and setup steps.
If you are running on limited GPU resources, use the quantised versions of the models. The model card on Hugging Face lists the available variants and their requirements.
SEA-LION is trained on and supports the national languages of Southeast Asia. This includes (among others) Indonesian, Thai, Vietnamese, Malay, Tamil, Khmer, Burmese, Lao and Filipino. For the authoritative and most current list, see the model documentation.
Yes. These are among the Southeast Asian languages SEA-LION is trained on. We are continually working with local and international partners to strengthen performance across all supported languages.
We are actively improving Tetum capability, working together with local and international partners. It is an area of ongoing development.
No. We evaluate and support the national languages of Southeast Asia. This does not include dialects such as Cantonese, Hokkien, Teochew, Hakka or other Singaporean dialects.
SEA-LION currently supports text and image (vision) inputs, and does not yet have a voice or audio model, though we do have plans to release text-to-speech (TTS).
Not yet. SEA-LION today supports text and vision, and does not currently have a voice or audio model. We do have plans to release text-to-speech (TTS). Follow our website and AI Singapore's social channels to hear when it is released.
SEA-LION and MERaLiON are complementary models that address different needs and work in tandem to strengthen Singapore's national AI foundation. SEA-LION, built by AI Singapore, is an open-source, SEA-relevant large language model. It is multilingual, multicultural and multimodal (handling text and vision), and is built for general tasks such as instruction following, tool use and reasoning. MERaLiON is a speech-to-text model that prioritises empathetic learning and understanding, starting with Singlish and Singapore's multicultural context and extending to regional languages. It also supports speech-related tasks such as emotion recognition and spoken dialogue summarisation. MERaLiON is developed by A*STAR's Institute for Infocomm Research (I²R).
The two leverage shared data, compute and research to build complementary text and speech models.
Our benchmark evaluations are published on the .
Make sure you are using the latest version of the SEA-HELM evaluation code base, which is updated as we add languages and metrics. We recommend running with the default SEA-HELM configuration.
We welcome collaboration across the public sector, private sector, academia and the non-profit space. The best first step is to submit the Collaborate form on or email us. To help us route your request quickly, please include a short overview of your organisation, your use case or proposed collaboration, and the languages or capabilities you are interested in.
We are always glad to hear from people who can help strengthen SEA-LION's coverage and quality. Share details of the data or capability you can offer, and we will let you know if it aligns with our current priorities. We do review fit case by case, so not every offer will match an active need at a given time.
AI Singapore/SEA-LION is a non-profit initiative funded by Singapore's National Research Foundation (NRF). We are not able to provide grants or funding for external projects. You are very welcome to use the open-source models freely for your work.
Reach out via the Collaborate form or email. Note that SEA-LION is already available on several major platforms, so do let us know what you have in mind and we can explore fit.
PAIP is a company-sponsored programme that helps organisations build applied AI capability within their teams. Because it is tailored to each organisation, we usually start with a short call to understand your team, your goals and the number of participants before sharing detailed materials. If you are interested, email us with your organisation and a brief description of your needs.
Please email us with details of your outlet, the focus of the story and your timeline. Media requests are handled together with AI Singapore's Marketing Communications team, and we will follow up to confirm.
We are glad to receive speaking and panel invitations. Email us with the event details, format, date and audience, and we will assess fit and timing.
If you have identified a potential security issue, please email us at with the details so we can route it to the right team.
Follow AI Singapore on LinkedIn and our other social channels, and keep an eye on the SEA-LION website. That is where we announce new model releases, events and opportunities.
Disruptive behavior: Deliberate disruption of discussions, events, or community activities is discouraged.
Launched in May 2017, AI Singapore is a national R&D programme in AI – the very first of its kind on a national level. Through SEA-LION, we aim to support under-represented population groups and low-resource languages in Southeast Asia. We welcome collaborations that further enhance our work and help our shared mission of democratizing AI.
We invite researchers, developers, language enthusiasts, and organizations to collaborate in enhancing and expanding the capabilities of our models. This page outlines the various ways you can work with us—whether it’s joint model building, research partnerships, code contributions, documentation improvements, or data sharing.
We appreciate your interest in contributing to SEA-LION, and we look forward to collaborating with you!
We offer various partnership modalities to suit different needs and interests, with a focus on open-sourcing results to benefit the wider community. For collaborations other than a self-serve model, an NDA/MOU/CA may have to be signed by both parties.
Using our model (self-service)
SEA-LION models are open-source so do feel free to download them for your own needs. If SEA-LION manages to be helpful to you/your team, we would love to hear about it and share your successes with our wider community. Send us a note at sealion@aisingapore.org.
Building models together, Data contributions
Have a unique dataset or domain expertise? Contribute data that we can incorporate into future SEA-LION releases. Collaborate with us to train specialized or improved versions of SEA-LION that will be open-sourced, benefiting both you and the broader community.
If there are areas of research topics where you think that AISG and yourself can work together to work on a common research agenda that can further the work of SEA-LION, please feel free to reach out to us as well. We will be open to joint publications arising from these collaborations too.
If you are interested in any form of collaboration, please feel free to reach out to sealion@aisingapore.org
In addition to formal partnerships, there are many individual and community-driven ways to help SEA-LION grow. Whether it’s fixing bugs, improving documentation, or proposing new features, as a developer, researcher, or just an enthusiast, your input matters.
Before you begin, please take a moment to review this guide, which outlines the contribution process, code of conduct, and how to get help if needed.
Before you start contributing, please ensure you have the following:
A GitHub account: If you don't have one, you can .
Familiarity with Git: You'll need to know the basics of Git for version control.
If you encounter any issues, bugs, or unexpected behavior while using SEA-LION, please help us by .
To report a bug:
Check if the issue has already been reported by searching the page.
If not, create a new issue with a descriptive title and detailed description of the problem you encountered.
Include relevant information such as your operating system, Python version, and any error messages.
We appreciate your suggestions for improving SEA-LION. If you have an idea for an enhancement or new feature, follow these steps:
Check if your suggestion has already been proposed in the section.
If not, create a new issue with a clear and concise title and a detailed description of your suggestion.
Include any relevant context or examples to illustrate the enhancement's value.
Please review and adhere to our . We expect all contributors and community members to treat each other with respect and kindness.
If you have questions, need assistance, or want to discuss contributions further, please feel free to contact us or open an issue for discussion.