Our Pledge

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.

Expected Behavior

To contribute to a positive and inclusive atmosphere, we expect all participants, including contributors, users, and maintainers, to:

1. Be respectful and considerate: Treat others with kindness, respect, and empathy. Recognize and embrace diversity in backgrounds, experiences, and opinions.

2. Be inclusive: Welcome and support people of all backgrounds, identities, and abilities. Avoid any form of discrimination or exclusionary behavior.

3. Listen actively: Pay attention to others' ideas, experiences, and feedback. Be open to constructive criticism and different points of view.

4. Show empathy: Understand that people may have different cultural norms, communication styles, and perspectives. Be patient and considerate when engaging with others.

5. 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.

6. 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:

1. Harassment: Any form of harassment, including but not limited to offensive comments, slurs, intimidation, or unwelcome advances, is strictly prohibited.

2. 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.

3. Hate speech: Hate speech, promoting violence, or advocating harm towards individuals or groups based on their identity is not allowed.

4. 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.

SEA-LION version 4.8, released in Sep 2026, is our latest collection of foundational, text generation optimized for Southeast Asia. Utilizing open dataset for continue pretraining and advanced post-training methodologies—such as agentic on-policy distillation delivers state-of-the-art regional fluency and high computational efficiency.

Built on the highly efficient NVIDIA Nemotron-3 Nano (30B-A3B) and Super (120B-A12B) base architectures, our models are optimized for precise text generation while maintaining low active-weight overhead.

• Refer to the detailed page to access model cards, usage scripts, and deployment configurations for the entire nemotron family:

  • Focused on parallel-only dataset for the core text-generation Foundations

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

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.

SEA-LION version 3.5, released in Apr 2025, is our first set of hybrid reasoning models trained on Southeast Asian data, each with their unique strengths:

• Llama-SEA-LION-v3.5-8B-R is suited for cost-efficient, low-latency applications where regional language support and advanced reasoning is critical.

• Llama-SEA-LION-v3.5-70B-R is suited for knowledge-intensive tasks, and for high-demand environments where advanced reasoning and comprehensive language comprehension are essential.

Trained off our SEA-LION v3 models, SEA-LION v3.5 is explicitly enhanced for reasoning tasks with the inclusion of thinking blocks, continuing our mission to create language models that understand and respond with greater cultural awareness and depth across Southeast Asia.

A distinctive feature of SEA-LION v3.5 is its dynamic reasoning toggle - Mode selection is managed through customizable chat template configurations, offering versatile functionality in handling both complex reasoning tasks and general text generation.

For detailed information of each of the SEA-LION v3.5 models, please refer to their individual documentation pages via the links above.

SEA-LION version 3, released in Dec 2024, is a collection of 3 models (and their variants), each with their unique strengths:

• Gemma-SEA-LION-v3-9B, based on Gemma2, is our best performing SEA-LION v3 model on SEA-HELM benchmarks for similar sized models

• Llama-SEA-LION-v3-8B, based on Llama 3.1 8B, now provides a larger context length of 128K

• Llama-SEA-LION-v3-70B, based on Llama 3.1 70B, is our largest model to date, also providing a 128K context length

With enhanced natural language reasoning (NLR) abilities and superior instruction-following in SEA languages, SEA-LION v3 sets new standards in multilingual AI for the region.

For detailed information of each of the SEA-LION v3 models, please refer to their individual documentation pages via the links above.

SEA-LION-Guard-v2 is our next-generation suite of action governance and safety models tailored specifically for the Southeast Asian AI ecosystem. Moving beyond standard binary content filtering, the v2 suite introduces real-time Action Governance, Bring Your Own Policy (BYOP) enforcement, and deep Multi-Turn Context Evaluation across 8 core regional languages (Burmese, English, Filipino, Indonesian, Malay, Tamil, Thai, and Vietnamese).

The v2 collection offers scalable parameters and context windows designed for diverse deployment requirements—from edge agent execution to enterprise cloud governance:

• (https://huggingface.co/collections/aisingapore/sea-lion-guard-v2) is a lightweight 31B, E4B and E2B parameter model with 256K context window for 31B and 128K context window for E4B and E2B. Optimized for ultra-low latency, it acts as an efficient edge guardrail for real-time tool interception and rapid API checks.

Compared to legacy moderation models, Gemma-SEA-LION-Guard-v2 shifts from freeform string generation to type-safe, calibrated decision probabilities across three fields:

SEA-LION version 4, released in August 2025, represents our first collection of multimodal models trained on Southeast Asian text. Each model offers unique strengths tailored to specific needs:

• A lightweight model optimized for mobile and edge devices, bringing robust SEA language understanding to environments with strict memory and latency constraints.

• Our powerful vision-language model, expertly trained to interpret both images and text with a deep, nuanced understanding of Southeast Asian cultural contexts.

• Our latest specialized vision-language models, featuring a native 256K context window and superior OCR capabilities for Indonesian, Thai, and Vietnamese. The 4B model

SEA-LION version 4.5, released in May 2026, agentic, and multimodal models optimized for Southeast Asia. Utilizing advanced post-training methodologies—such as knowledge distillation and model merging—this suite delivers state-of-the-art regional fluency, precise tool use, and high computational efficiency.

Our highly efficient 4-billion parameter series built on Gemma 4. It is optimized for precise function-calling, structured JSON outputs, and autonomous agentic tool-use with minimal memory overhead.

• Refer to the detailed page to access model cards, usage scripts, and deployment configurations for the entire Gemma family:

  • Core Agentic Foundations

Our high-capacity flagship causal and vision-language series built on Qwen3.6. It features a native 262K context window, robust multi-turn reasoning, and repository-level coding adapted for regional linguistic and cultural contexts.

List of SEA-LION research papers published:

, released in March 2026, is a suite of high-performance encoder models specifically architected for Southeast Asian languages. The collection features two primary product lines:

• : Built on the ModernBERT architecture, these models feature a native 8,192 token context window and are optimized for high-throughput production environments and long-document RAG.

• : Fine-tuned from the E5-large foundation, these models are designed for maximum semantic accuracy in retrieval and similarity tasks across the region's diverse linguistic landscape.

For detailed information on specific models and checkpoints, please refer to their individual documentation pages.

SEA-Guard, released on 4 Feb 2026, is our premier suite of safety and moderation models tailored for the Southeast Asian landscape. This collection focuses on robust visual and textual guardrails, ensuring AI deployments remain secure and culturally compliant across the region.

The SEA-Guard Collection

The SEA-Guard suite currently consists of 4 specialized models:

• Qwen-SEA-Guard-4B (https://huggingface.co/aisingapore/Qwen-SEA-Guard-4B-040226) is a 4B Image-to-Text model serving as a lightweight visual safety guardrail, specifically optimized for efficient edge applications where low latency is critical.

• Qwen-SEA-Guard-8B (https://huggingface.co/aisingapore/Qwen-SEA-Guard-8B-040226) is an 8B Image-to-Text model that offers a balanced visual moderation solution with stronger reasoning capabilities for more nuanced content detection.

• (https://huggingface.co/aisingapore/Llama-SEA-Guard-8B-040226) is an 8B Text Generation model. This text-only safety specialist is optimized for chat moderation and policy enforcement, ensuring safe interactions in dialogue systems.

• (https://huggingface.co/aisingapore/Gemma-SEA-Guard-12B-040226) is a high-capacity 12B Image-Text-to-Text multimodal safety model designed for complex content analysis, capable of interpreting intricate relationships between visual and textual data.

The SEA-Guard collection advances our mission to build AI that is not only culturally intelligent but also inherently safe. By providing specialized guardrails for both text and vision, these models enable developers to deploy AI solutions across Southeast Asia with confidence in their safety and compliance standards.

Note: For detailed integration guides and benchmark data for each SEA-Guard model, please refer to their individual documentation pages via the links above.

This section covers the various environments where SEA-LION models can be hosted, ranging from fully managed cloud services to self-hosted high-performance infrastructure. SEA-LION models are freely available for download. The following guides provide technical how-tos for setting up SEA-LION inference using different approaches:

1. Using our provided SEA-LION API

2. Running SEA-LION on a local machine (coming soon)

3. Deploying SEA-LION on the cloud

4. Leveraging our Partner API Platforms

SEA-LION models can be used as intelligent agents capable of tool-calling, multi-step reasoning, and complex task execution. The SEA-LION v3 family, such as aisingapore/Llama-SEA-LION-v3-70B-IT, can be utilized for agent-based applications with tool-calling capabilities.

Agents built with SEA-LION can:

• Execute complex, multi-step workflows

• Integrate with external tools and APIs

• Perform real-time information retrieval

• Handle specialized tasks like translation, calculation, and web search

• Chain multiple operations together intelligently

The following guides provide technical how-tos for building SEA-LION agents using different frameworks:

Learn how to build powerful agents using AWS's Strands Agents SDK, including:

• Basic agent setup using in-built tools

• Custom tool development

• Using agents as tools

• Multi-tool agent example

Develop and deploy agents with Google ADK:

• Basic agent setup with custom tools and agent-as-a-tool

• Custom tool development

• Using agents as tools

• Running multi-tool agent via in-built web interface/CLI

For optimal agent performance, we recommend using the following SEA-LION models:

• aisingapore/Llama-SEA-LION-v3-70B-IT/aisingapore/Llama-SEA-LION-v3-8B-IT - Suitable for tool-calling, long context tasks or creative response.

• aisingapore/Gemma-SEA-LION-v3-9B-IT - Suitable for smaller tasks with succinct response.

• aisingapore/Llama-SEA-LION-v3.5-70B-R

Tool Integration: SEA-LION agents can seamlessly integrate with both built-in tools (calculations, time queries) and custom tools (web search, translation, APIs).

Multi-Agent Systems: Agents can be composed together in multiple ways. Specialized agents serving as tools, running in sequential workflow, or passing to each other in a loop.

Streaming Responses: Built-in support for real-time streaming responses for better user experience.

Flexible Configuration: Easy environment-based configuration for different deployment scenarios.

SEA-LION Model Context Protocol (MCP) Server provides a simple way to access SEA-LION’s multilingual models and Southeast Asian language tools from MCP-compatible clients.

The Model Context Protocol (MCP) is an open standard for connecting AI models to external tools, data sources, and services. It defines a common interface so that clients (such as IDEs, chat assistants, or AI agents) can discover and use capabilities provided by servers without custom integrations for each tool. Powered by the SEA-LION API, the MCP server exposes capabilities including translation and localisation, and safety classification.

Usage of the SEA-LION API is subject to our and

Claude Desktop handles the OAuth flow for you automatically. It's the most straightforward way to connect.

1. Open Claude Desktop and go to Settings → Connectors → Customize

2. Click Add Custom Connector and select Web

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.

Key Features of SEA-LION

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:

1. How SEA-LION compares to other available models along different metrics

2. What SEA-HELM is and the four key capabilities it is evaluated on: English performance, Proficiency in SEA chat, Instruction-following and Linguistic tasks

3. 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:

1. Pre-Training data

2. Model training code

3. Fine-Tuning data

4. 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 .

The SEA-LION models have served as a foundation for developing localized AI solutions tailored to specific linguistic and cultural needs. Over multiple iterations, other regions have built upon SEA-LION’s architecture to create specialized versions that enhance language understanding for their respective regions. These models leverage SEA-LION’s robust multilingual capabilities while being fine-tuned with localized datasets, ensuring better performance in regional contexts.

SEA-LION Model Tree

By leveraging SEA-LION’s architecture, these localized models provide AI solutions that align more closely with native language requirements. As SEA-LION continues to evolve, more localized versions are expected to emerge, further expanding the reach and effectiveness of AI in Southeast Asian languages.

Last updated: 2025-10-17

SEA-VLM is an instruct-tuned vision-text model for the Southeast Asia (SEA) region.

Gemma-SEA-LION-v4-27B-VL has undergone post-training using instruction-image pairs datasets in Burmese, English, Indonesian, Khmer, Lao, Malay, Mandarin, Tagalog, Tamil, Thai and Vietnamese, comprising approximately 540k samples in total, to create Gemma-SEA-LION-v4-27B-VL.

Gemma-SEA-LION-v4-27B-VL inherits Gemma 3's:

• Large 128K context length

• Image and text understanding capabilities, including document comprehension, visual Q&A, and image-grounded reasoning

• Advanced function calling and structured outputs to allow for seamless integration into larger systems

We performed post-training in English and SEA languages on Gemma-SEA-LION-v4-27B-IT, a decoder model using the Gemma 3 architecture, to create Gemma-SEA-LION-v4-27B-VL.

For tokenization, the model employs the default tokenizer used in Gemma 3 27B IT.

• Developed by: SEACrowd and Products Pillar, AI Singapore

• Funded by: Singapore NRF

• Shared by: SEACrowd and Products Pillar, AI Singapore

• Model type: Decoder

As of 15 October 2025, Gemma-SEA-LION-v4-27B-VL excels at Southeast Asian (SEA) tasks when compared to other open models with fewer than 200 billion parameters and demonstrates performance comparable to that of larger and top closed models. For detailed rankings, please refer to the .

The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes.

The model was not tested for robustness against adversarial prompting. It is important for users to be aware that our model exhibits certain limitations that warrant consideration. Like many LLMs, the model can hallucinate and occasionally generates irrelevant content, introducing fictional elements that are not grounded in the provided context. Users should also exercise caution in interpreting and validating the model's responses due to the potential inconsistencies.

Limitations

In terms of text capability, Gemma-SEA-LION-v4-27B-VL has been trained and fine-tuned exclusively on the vision-text backend. As a result, its text capabilities are expected to be comparable to those of Gemma-SEA-LION-v4-27B-IT, and may not exhibit significant improvements or differences in this area.

Use the code below to get started with the model using the 🤗 Transformers library.

The dataset comprises vision-text paired in Burmese, English, Indonesian, Khmer, Lao, Malay, Mandarin, Tagalog, Tamil, Thai and Vietnamese languages, collected from a mixture of sources including web data, code, open-source datasets.

• Training regime:

We perform SFT using 540k of vision-text samples written in 10 languages. Then, we perform model merging with Gemma3-27B-IT to preserve general vision-text knowledge.

For details on Gemma-SEA-LION-v4-27B-VL performance, please refer to the SEA-LION.ai blogpost, .

Carbon emissions can be estimated using the presented in .

• Hardware Type: Nvidia H200 140GB GPUs

• Hours used: 13 hrs

• Cloud Provider: SMC H200

• Compute Region: Singapore

This is the repository for the commercial instruction-tuned model. The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes.

AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of the National Research Foundation or the National University of Singapore.

SEA-HELM Audio is a Southeast Asian (SEA) multilingual evaluation framework that extends the text-based SEA-HELM benchmark. It is designed to assess performance degradation in cascaded automatic speech recognition and large language model (ASR–LLM) pipelines.

SEA-HELM Audio serves as a valuable resource for developers looking to deploy and evaluate voice agents, providing empirical baselines for how linguistic accuracy and instruction-following hold up when transitioning from text to spoken interactions. It currently covers English, Indonesian, Malay, Tamil, and Chinese.

Motivations

Limitations of Text-Only Benchmarks for Voice Agents

While Large Language Models (LLMs) continue to achieve high benchmark scores on written text, deploying these models within real-world voice pipelines introduces acoustic transcription errors, phonetic ambiguities, and domain-specific degradation. Existing spoken language benchmarks do not adequately capture Southeast Asian linguistic nuances. There was a critical need to systematically measure the impact of speech input on downstream LLM task performance across SEA languages.

Quantifying the "Speech Gap"

To evaluate how voice interfaces bottleneck LLM reasoning, SEA-HELM Audio establishes a controlled framework to quantify the Speech Gap (Δ = Score_text − Score_speech). This absolute difference represents the "modality penalty," quantifying precisely how much task performance degrades due to transcription errors or missing acoustic signals.

Core Pillars of SEA-HELM Audio

1. Spoken Prompt Construction

Because the original SEA-HELM contains only written prompts, SEA-HELM Audio natively creates spoken input. Each prompt is read aloud by a Text-to-Speech (TTS) synthesizer that plays the role of the user. To ensure authenticity, only the content a user would actually say is synthesized, and written conventions (like multiple-choice markers) are adapted so they read naturally aloud in each specific language.

2. Dual-Path Evaluation Flow

To isolate the exact cost of the speech front-end, each SEA-HELM prompt is evaluated in two parallel ways:

• The Speech Path: The text prompt is converted to speech via TTS, transcribed back to text by an ASR system, and the resulting transcript is answered by the LLM.

• The Text Path: The LLM answers the original, ground-truth text prompt directly.

Because both paths share the same LLM and rule-based verifier, any difference in performance isolates the failure of the acoustic pipeline.

The framework accommodates diverse pipeline architectures, evaluating both open-weight multilingual baselines (e.g., Qwen3-ASR, MERaLiON3-ASR) and commercial API services (e.g., ElevenLabs). It measures transcription quality (Word Error Rate and Character Error Rate) to separate how much the ASR changed the prompt from how much those changes actually derailed the downstream LLM task.

All responses are scored with rule-based verifiers taken from SEA-HELM, keeping evaluations deterministic and identical across paths. Initial benchmarking focuses on highly verifiable subsets like SEA-HELM IFEval (instruction following) and Global-MMLU-Lite (multiple-choice reasoning).

The leaderboard allows users to compare different ASR-LLM configurations, highlighting the performance penalty of the speech modality. Initial empirical results have exposed severe degradation across the board. For example, while models achieve exceptionally high scores on text-only instruction following, end-to-end speech accuracy drops significantly.

Notably, the benchmark reveals a profound "Tamil Bottleneck," where the modality penalty exceeds 40 percentage points regardless of whether open-weight or commercial ASR systems are used, pointing to a severe acoustic data scarcity in regional speech modeling.

SEA-HELM Audio forms part of the broader ATLAS-Eval ecosystem. Our methodology, evaluation frameworks, and related research papers (including the formal SEA-HELM Audio: Evaluating Multilingual ASR-LLM Pipeline Performance preprint) are publicly accessible.

While SEA-HELM Audio establishes a solid baseline using synthesized speech to measure the "floor" of performance degradation, involve natural accents, background noise, and conversational code-switching. Future iterations of SEA-HELM Audio aim to incorporate naturalistic, crowdsourced audio corpora to provide an even more robust and comprehensive evaluation of Southeast Asian voice AI.

For more details and information on SEA-HELM Audio, you can refer to the following materials and platforms:

1. SEA-HELM Audio: Evaluating Multilingual ASR-LLM Pipeline Performance

2. SEA-HELM Audio Benchmark Dataset Available on Hugging Face:

3. SEA-HELM Audio Evaluation Repository Available on GitHub:

4. SEA Audio Leaderboard & Voice-Agent Playground Interactive evaluation platform at

In the previous guide, we have shown how to leverage on Amazon Bedrock‘s Custom Model Import to deploy SEA-LION in the cloud. After importing the SEA-LION models, you can build applications with the AWS SDK. This guide describes an alternative method to build the applications with OpenAI-compatible APIs served by the Bedrock Access Gateway.

Prerequisites

The SEA-LION model is imported and available on Amazon Bedrock. For imported models, you are charged for model inference. Please refer to the Amazon Bedrock Pricing page for the latest information.

Please check that the following are installed on your development machine.

• Git

• Python

Local Installation

If you have already cloned the repository from the previous , skip to the next step. The repository is a fork to support the imported models. At the time of writing, the original repository supports only.

Navigate to the src directory.

Check the and environment variables. Please ensure that AWS_REGION is set to the region where the model is imported.

• AWS_ACCESS_KEY_ID

• AWS_SECRET_ACCESS_KEY

• AWS_SESSION_TOKEN

• AWS_REGION

Before running the gateway, it is a good practice to create a virtual environment to isolate the app. Please follow these steps to create a virtual environment, or feel free to use your preferred tool.

Initialise the virtual environment.

Activate the virtual environment.

Install the packages.

Start the gateway.

Alternatively, start the gateway with Docker Compose if it is installed.

List the models with the following curl script. The imported model is in the list with the format arn:aws:bedrock:<AWS_REGION>:<ACCOUNT_ID>:imported-model/<MODEL_ID>

Test the chat completion API with the following curl script. Replace <AWS_REGION>, <ACCOUNT_ID> and <MODEL_ID> with the values from the imported model’s ARN (Amazon Resource Name) from the model list.

With the gateway, the demo from the can use the to integrate with the imported model. Follow the steps in the guide to set it up. Run the demo with the --api parameter.

• SEA-LION Models on Hugging Face:

• Bedrock Access Gateway:

  • Production deployment:

Last updated: 2026-02-05

SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.

Gemma-SEA-LION-v4-4B-VL is a 4-billion parameter Vision-Language Model (VLM) built upon the gemma-3-4b-it architecture. To ensure domain adaptation for the region, the model underwent rigorous post-training on a curated dataset of approximately 8.54 million instruction-text pairs. This curated dataset included various SEA Languages such as Burmese, Malay, Tagalog and Tamil. This extensive post-training instills multilingual and multicultural fluency for both SEA languages and English.

Gemma-SEA-LION-v4-4B-VL inherits the image and text capabilities from gemma-3-4b-it alongside its large context length of 128K tokens. Additionally, beyond extending the multilingual capabilities of the original gemma model for SEA languages, we experimented with:

1. Adding function calling to the model to allow for this model to be used in tool calling applications

Common questions we receive at the SEA-LION inbox, grouped by topic. If your question is not covered here, write to us at or use the Collaborate form on .

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 .

• Chat with SEA-LION on our .

SEA-Guard is a collection of safety-focused Large Language Models (LLMs) built upon the SEA-LION family, designed specifically for the Southeast Asia (SEA) region.

SEA-LION stands for Southeast Asian Languages In One Network and is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.

This model is a fine-tuned version of , and on 1M instruction-following pairs. For more details on training data, please refer to the paper .

For tokenization, the model employs the default tokenizer used in Qwen3-VL.

• Developed by: AI Products Pillar, AI Singapore

• Funded by:

SEA-Guard is a collection of safety-focused Large Language Models (LLMs) designed specifically for the Southeast Asia (SEA) region. While the collection comprises four distinct models, we currently offer a single API endpoint that exclusively serves the . You can generate an API key to access this model at .

SEA-LION stands for Southeast Asian Languages In One Network and is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.

This model is a fine-tuned version of on 1M instruction-following pairs. For more details on training data, please refer to the paper .

For tokenization, the model employs the default tokenizer used in Gemma 3.

• Developed by: AI Products Pillar, AI Singapore

The following guide explains the procedures for deploying a SEA-LION model on a Linux server.

• OS: Linux

• Python: 3.9-3.12

• vLLM version: 0.10.1.1

• uv 0.7.x installed

Our Gemma-SEA-LION-v4-27B-IT model is now supported on the Cloudflare Workers AI for on-demand inferencing, subject to Cloudflare's and .

This quickstart guide will instruct you through setting up your Workers AI account and REST API to easily experiment with Gemma-SEA-LION-v4-27B-IT. For more detailed information, you can refer directly to the latest guide on Cloudflare

Sign up for a if you have not already done so.

You need your API token and Account ID to use the REST API.

To get these values:

• In the Cloudflare dashboard, go to the Workers AI page and select REST API:

SEA-HELM (Southeast Asian Holistic Evaluation of Language Models) is a comprehensive benchmark suite designed to evaluate the linguistic and cultural competencies of Large Language Models (LLMs) in Southeast Asian (SEA) languages.

Formerly known as BHASA, SEA-HELM has been expanded and integrated with HELM to provide a more rigorous and authentic evaluation. It addresses the critical need for multilingual and multicultural benchmarks in the rapidly evolving field of LLMs.

SEA-HELM serves as a valuable resource for developers looking to create and assess LLMs for linguistic accuracy and cultural sensitivity in the Southeast Asian context. It currently covers Filipino, Indonesian, Javanese, Sundanese, Tamil, Thai, and Vietnamese.

Motivations

Several key factors motivated our development of SEA-HELM:

Lack of Comprehensive SEA Language Benchmarks

Existing LLM benchmarks are capable of evaluating specific capabilities of LLMs in English and some mid- to low-resource languages, including those in the Southeast Asian (SEA) region. However, there has not been a comprehensive and authentic evaluation suite developed specifically for SEA languages, a problem exacerbated by the lack of both training and testing data on the internet.

Need for Multilingual and Multicultural Evaluations

There's an increasing need for benchmarks that assess not only linguistic capabilities but also cultural representation and sensitivity. Evaluations of linguistic and cultural representation are essential for gauging the efficacy and fairness of language models.

SEA-HELM deliberately incorporates community participation by involving native speakers to ensure linguistic accuracy and cultural authenticity. For example, the evaluation suite includes a cultural evaluation dataset for Filipino developed in collaboration with community members from the Philippines.

Core Pillars of SEA-HELM

SEA-HELM is organized into five core evaluation pillars, designed to comprehensively assess various competencies of LLMs.

Each evaluation task under SEA-HELM uses prompts in the target language to ensure that the LLM can interpret native instructions correctly. The datasets used are either originally written in the native language or carefully translated by native speakers to avoid translation errors. This ensures that the evaluation is authentic and relevant to the linguistic nuances of each language.

This pillar focuses on evaluating the fundamental natural language processing abilities of LLMs in Southeast Asian languages, such as language understanding, generation, and reasoning.

NLU tasks assesses the model's ability to comprehend text. The tasks included in this competency are question answering and sentiment analysis.

• Question Answering (QA) evaluates the model's ability to understand a question and extract the answer from a given text. SEA-HELM uses datasets like TyDi QA for Indonesian, XQuAD for Vietnamese and Thai, and IndicQA for Tamil. The models are prompted to answer questions by extracting the answer from a provided paragraph.

• Sentiment Analysis: determines sentiment expressed in a text. SEA-HELM uses datasets like NusaX for Indonesian, UIT-VSFC for Vietnamese, Wisesight Sentiment for Thai, and IndicSentiment for Tamil. The models are prompted to determine the sentiment of a given sentence and respond with a single word: Positive, Negative, or Neutral.

NLG tasks evaluates a model's ability to generate human-like text. The tasks included in this competency are machine translation and abstractive summarization.

• Machine Translation assesses the model's ability to translate text from one language to another. SEA-HELM evaluates translation between English and target SEA languages. The models are prompted to translate a given text into a specified language.

• Abstractive Summarization requires the model to read a document, identify the key points, and summarize them into a coherent and fluent text. SEA-HELM uses the XLSum dataset for all four target languages. The models are prompted to summarize an article in one or two sentences in the specified language

NLR tasks assesses the model's ability to reason and draw inferences from text. The tasks included in this competency are causal reasoning and Natural Language Inference (NLI).

NLI tasks involve determining whether a given premise entails or contradicts a hypothesis. SEA-HELM uses datasets like IndoNLI for Indonesian, XNLI for Vietnamese and Thai, and IndicXNLI for Tamil.

This pillar focuses on evaluating capabilities unique to LLMs, such as instruction following and chat capabilities.

• Instruction following assesses the ability of LLMs to follow human instructions and adhere to specified formats using SEA-IFEval, a benchmark we manually created collaboratively with native speakers.

• Chat capability evaluates the ability of LLMs to engage in human-like conversations using SEA-MTBench, another manually translated and localised version of the popular MT-Bench dataset.

This pillar focuses on systematically diagnosing language proficiency and grammatical understanding of language models in Southeast Asian languages by utilizing LINDSEA (LINguistic Diagnostics for Southeast Asian languages).

LINDSEA is a high-quality, manually-crafted linguistic dataset designed to provide a fine-grained evaluation of a model’s linguistic abilities, and is the first dataset of its kind created for SEA languages.

This pillar is designed to assess cultural representation and sensitivity in language models. This pillar recognizes the importance of evaluating LLMs on their understanding and appropriate handling of cultural nuances, social norms, and values specific to Southeast Asian cultures.

SEA-HELM uses cultural diagnostics to probe for both cultural representation and sensitivity. The goal is to ensure that LLMs used in SEA contexts are not only linguistically accurate but also culturally aware and respectful.

SEA-HELM achieves authentic cultural representation through a strong participatory approach that includes native speaker communities. For example, SEA-HELM includes a cultural evaluation dataset for Filipino developed in collaboration with community members from the Philippines, which resulted in KALAHI.

This pillar ensures that language models (LLMs) do not produce harmful or unsafe outputs, especially in the context of Southeast Asian (SEA) languages and cultures. It recognizes that multilingual inputs, particularly in lower-resource languages common in the SEA region, can increase the likelihood of LLMs generating unsafe responses. The pillar aims to protect users interacting with these models in SEA languages from issues like hate speech.

Currently, the Safety pillar focuses on toxicity detection as its primary task, with coverage for Indonesian, Thai, Vietnamese, and Filipino. This involves identifying toxic content such as hate speech and abusive language in text, which is crucial for content moderation. SEA-HELM uses datasets like MLHSD for Indonesian, Thai Toxicity Tweet for Thai, ViHSD for Vietnamese, and PH Elections Toxicity for Filipino.

While passing these evaluations is a necessary step, it is not a guarantee of complete safety in real-world scenarios, as it is impossible to cover every type of unsafe response. SEA-HELM plans to expand this pillar to include a broader range of safety-related tasks in the future.

To provide transparency, comparative insights and understanding of models' multilingual and multicultural performance, SEA-HELM maintains a .

The offers multiple views, including overall scores, language-specific scores, and detailed task scores, allowing users to delve deeper into the evaluation results.

The full codebase for SEA-HELM is provided open-source on .

Our SEA-HELM evaluation datasets are publicly available on .

While SEA-HELM aims for holistic evaluations, it is not yet exhaustive in its coverage of languages and tasks. Areas for improvement include:

• Expanding SEA language coverage beyond the current seven languages (e.g., Burmese, Khmer, Lao).

• Exploring automatic LLM evaluations for better benchmarking.

• Enhancing the safety pillar to include more real-world risk factors.

• Adding more cultural and linguistic assessments

For more details and information on SEA-HELM, you can refer to the following materials:

A step-by-step guide to deploying AI Singapore's SEA-LION models to run as an endpoint on Google's Vertex AI.

1. Go to Google's Model Garden

2. Scroll down (in the main page) to the "All Partners" section, and click on "Hugging Face".

3. Set the “Filter by Name” to “aisingapore”. A list of SEA-LION models will appear.

1. Click the model you wish to deploy, and a side panel will pop up. Here you can set the name of your endpoint, select the region and machine spec to deploy the model to. Then click on Deploy to complete the deployment.

   (Note: make sure you have enough hardware quota in your selected region)

1. The deployment takes some time (depending on the size of the model). When done, you will see the deployed endpoint in your

   Take note of your endpoint ID.

1. To test your deployed endpoint using a Python program,

   a. Get a Vertex AI service key from your administrator, or create one yourself if you have the right permissions on GCP.

   b. Install the for your system.

   c. Make sure your Google Cloud SDK is properly installed and authenticated by running the following command: gcloud auth login

   d. Create a new Python environment and install the following packages: pip install google-cloud-aiplatform openai

2. Create a Python program vt-test.py containing the following code:

1. Make sure your Vertex AI service key JSON file (named vt-svc-key.json) is in the same directory as your Python program.

2. Run your program python vt-test.py and see the result:

1. Below is a sample bash script curl_test.sh for calling the endpoint via CLI using curl:

When run,

The SEA-LION models are open source and freely available for research and commercial use. A common question we receive from developers is how to host and configure SEA-LION for model inference in their own environments. When it comes to deploying our models, organizations have several hosting options available to them. One such approach involves using the Custom Model Import feature in Amazon Bedrock, which allows for integration within Amazon’s cloud infrastructure.

At the time of writing, the supported model architectures include Llama 3 and Llama 3.1. The Llama-SEA-LION-v3-8B-IT and Llama-SEA-LION-v3-70B-IT models are supported. The SEA-LION models not built using the Llama architecture (such as Gemma2 and MPT) are not supported.

This article outlines the process of importing the SEA-LION model into Amazon Bedrock and includes a demo application that integrates with the imported model.

Prerequisites

If you do not already have an Amazon Web Services (AWS) account, please sign up first.

Note that in following this guide, the following AWS services are paid services and will incur costs:

• Amazon Bedrock

• Amazon S3

Please check that the following are installed on your development machine.

is a fully managed service that simplifies the deployment and scaling of AI models. It provides access to high-performing foundation models, enabling a serverless experience for model deployment and integration. With Custom Model Import, users can upload their own models and use a unified platform for AI development.

There is no charge to import a custom model to Bedrock. Once you import a model, you will be able to access it on-demand without requiring to perform any control plane action.

You are only charged for model inference, based on the number of copies of your custom model required to service your inference volume and the duration each model copy is active, billed in 5-minute windows.

A model copy is a single instance of an imported model ready to serve inference requests. The price per model copy per minute depends on factors such as architecture, context length, AWS Region, compute unit version (hardware generation), and is tiered by model copy size.

A monthly storage cost per Custom Model Unit is applicable. The Custom Model Units needed to host a model depend on a variety of factors — notably the model architecture, model parameter count, and context length. The exact number of Custom Model Units needed will be determined at the time of import.

Please refer to the page, including the Custom Model Import pricing under the Pricing Details section, for the latest information.

The model used in this guide is . This section describes the steps to import the model.

Upload the contents of to an Amazon S3 bucket. Please take note of the .

As the time of writing, Amazon Bedrock Custom Model Import is supported in the us-east-1 (N. Virginia) and us-west-2 (Oregon) regions. Please check the for the latest information.

Select a supported region in the AWS Console. Navigate to Amazon Bedrock, Imported Models. Click the Import Model button.

Input the model name and the S3 location of the uploaded model. Update the other accordingly. Click Import Model to start the import.

After the import is completed, locate the model in Imported models.

Click the copy button to copy the ARN (Amazon Resource Name).

The demo application uses the to integrate with the imported model.

Clone the . The repository is a fork to include a functional demo and to support the imported models via the gateway.

Navigate to the demo directory.

Copy the environment file.

Edit the value of ENDPOINT_ARN in the environment file and paste the ARN of the imported model.

Edit the value of AWS_REGION (e.g. us-east-1) in the environment file to match the region where the model was imported.

Before running the demo, it is a good practice to create a virtual environment to isolate the app. Please follow these steps to create a virtual environment, or feel free to use your preferred tool.

Initialise the virtual environment.

Activate the virtual environment.

Install the packages.

Set up the Boto3 credentials:

Run the demo.

Amazon Bedrock Custom Model Import optimizes the hardware utilization by removing the models that are not active. The demo might throw an exception that indicates the model is not ready for inference. Please refer to and customize your applications to handle it gracefully.

The demo uses the to integrate with the imported model. If you are looking for how to work with OpenAI-compatible APIs, given their popularity, please refer to the next .

• SEA-LION models on Hugging Face:

• Import a customized model into Amazon Bedrock:

Llama-SEA-LION-v3.5-8B-R is a hybrid model offering versatile functionality, handling both complex reasoning tasks and general text generation, with mode selection managed through the tokenizer's chat template.

We performed instruction tuning in English and also in SEA languages such as Filipino, Indonesian, Tamil, Thai and Vietnamese on our , a decoder model using the Llama 3.1 architecture, to create Llama-SEA-LION-v3.5-8B-R.

By leveraging on SEA-LION v3’s strong foundation, Llama-SEA-LION-v3.5-8B-R ensures broader accessibility and usability, empowering diverse communities and use cases throughout the region. It is particularly suited for cost-efficient, low-latency applications where regional language support and advanced reasoning is critical.

For tokenisation, the model employs the default tokenizer used in Llama 3.1 70B Instruct. The model has a context length of 128k.

At a glance:

• Model type: Decoder

Llama-SEA-LION-v3.5-70B-R is a hybrid model offering versatile functionality, handling both complex reasoning tasks and general text generation, with mode selection managed through the tokenizer's chat template.

We performed instruction tuning in English and also in SEA languages such as Filipino, Indonesian, Tamil, Thai and Vietnamese on our , a decoder model using the Llama 3.1 architecture, to create Llama-SEA-LION-v3.5-70B-R.

By leveraging on SEA-LION v3’s strong foundation, Llama-SEA-LION-v3.5-70B-R ensures broader accessibility and usability, empowering diverse communities and use cases throughout the region. It is particularly suited for knowledge-intensive tasks, and for high-demand environments where advanced reasoning and comprehensive language comprehension are essential.

For tokenisation, the model employs the default tokenizer used in Llama 3.1 70B Instruct. The model has a context length of 128k.

At a glance:

• Model type: Decoder

Last update: 2026-05-19

SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.

Gemma-SEA-LION-v4.5-E2B-IT built upon the gemma-4-E2B-it architecture with 2.3B effective (5.1B with embeddings). To ensure deep domain adaptation, the model underwent distillation from google/gemma-4-31B-it on an updated aisingapore/SEA-Instruct-2602, instilling multilingual and multicultural fluency across English and key SEA languages: Burmese, Indonesian, Filipino (Tagalog), Malay, Tamil, Thai, and Vietnamese.

Gemma-SEA-LION-v4.5-E2B-IT inherits the following features from Gemma 4:

• Native System Prompt Support – Gemma 4 introduces native support for the system role, enabling more structured and controllable conversations.

Last update: 2025-12-1

SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.

Qwen-SEA-LION-v4-VL contains 4/8-billion parameter Vision-Language Models (VLM) built upon the Qwen3-VL-4B/8B-Instruct architecture. To ensure domain adaptation for the region, the model underwent rigorous supervised fine-tuning (SFT) on a curated dataset of approximately 9 million instruction-text pairs. This extensive post-training instills multilingual and multicultural fluency, covering English and 7 key SEA languages: Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese.

Qwen-SEA-LION-v4-4B/8B-VL inherits the following features from Qwen3-VL:

• Long-Context Multimodal Architecture (Native 256K context window)

Last updated: 2025-10-17

SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.

Qwen-SEA-LION-v4-32B-IT is based on Qwen3, which provides a strong foundation with support for over 100 languages and advanced reasoning capabilities. The model underwent continued pre-training on approximately 100B tokens sampled from the SEA-Pile v2 pretraining corpus of over one trillion tokens across 7 SEA languages: Burmese, Indonesian, Malay, Filipino, Tamil, Thai, and Vietnamese. Finally, it was post-trained on a high-quality dataset of approximately 8 million question-and-answer pairs to create the final instruction-tuned model.

Qwen-SEA-LION-v4-32B-IT inherits the following features from Qwen3-32B:

• 32,768 of context length natively

SEA-LION stands for Southeast Asian Languages In One Network.

We performed continued pre-training in English and SEA languages on Qwen3-32B, a decoder model using the Gemma 3 architecture, and post-training to create Qwen-SEA-LION-v4-32B-IT.

Last updated: 2026-02-05

SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.

SEA-LION stands for Southeast Asian Languages In One Network.

Apertus-SEA-LION-v4-8B-IT is a 8-billion parameter model built upon the Apertus-8B-Instruct architecture. To ensure domain adaptation for the region, the model underwent rigorous post-training on a curated dataset of approximately 8.54 million instruction-text pairs.

This extensive post-training instills multilingual and multicultural fluency, covering key SEA languages such as Burmese, Malay, Tagalog and Tamil. The dataset also includes 347,000 tool-calling instruction-text pairs to impart these capabilities, in addition to linguistic fluency.

Apertus-SEA-LION-v4-8B-IT is designed as a fully open model; to align with this core philosophy, we have released the datasets used for post-training, as well as the evaluation codes and datasets used to evaluate the model.

These resources can be accessed via the link below.

Last update: 2026-03-16

SEA-LION is a collection of Large Language Models (LLMs) and encoders which have been pretrained and fine-tuned for the Southeast Asia (SEA) region.

SEA-LION stands for Southeast Asian Languages In One Network.

The SEA-LION-Embedding-E5-600M model is a Sentence Transformer optimised for 11 Southeast Asian languages. It has been fine-tuned from the base, mapping sentences and paragraphs to a 1024-dimensional dense vector space. This model is designed for high-accuracy semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and RAG (Retrieval-Augmented Generation) workflows. It leverages the robust XLM-RoBERTa architecture pretrained on 100 languages, optimised here for 11 Southeast Asian languages: Burmese, Chinese, English, Filipino, Indonesian, Khmer, Lao, Malay, Tamil, Thai, and Vietnamese.

The SEA-LION-E5-Embedding-600M model is a Sentence Transformer built on the which was initialised from architecture.

• Model Type: Sentence Transformer