Catalysing AI Innovation for
Southeast Asia

SEA-LION is a family of efficient, open-source, multilingual, multimodal language models designed to understand Southeast Asia’s diverse languages, cultures, and contexts.

SEA-LION Telegram SEA-LION Playground
HomepageGlobeImage SEA-LION AI

Latest Announcements

SEA-LION Insights

Stay up to date with the latest developments from SEA-LION. Our blog features updates on model releases, research breakthroughs, datasets, benchmarks, and stories from the growing Southeast Asian AI community.

Recently published:

SEA-LION v4.5

We are expanding our SEA-LION v4.5 family to include state-of-the-art models that work well as agents with understanding of SEA languages, cultures, and local knowledge. We have also trained our own speculative decoder to turbo charge our large Qwen model for up to 6X greater efficiency.

Find Out More

SEA-Guard

Introducing our new collection of safety-focused Large Language Models (LLMs) built upon the SEA-LION family. SEA-Guard models are specifically fine-tuned to detect, moderate, and handle content according to Southeast Asian cultural norms and safety standards. From lightweight visual safety models for edge deployment to high-capacity multimodal models for complex content analysis—there’s a SEA-Guard model for every safety need.

Find Out More

Model Catalogue

Explore our family of models and see our latest release.

Explore Model

Project ATLAS

Join our global community-driven open platform to help builders make AI more representative, more locally useful, and safer for the communities. Contribute and explore datasets powering future models.

Explore ATLAS

SEA-HELM Leaderboard

Track model performance and gain insights with SEA-HELM, assessing large language models across Southeast Asian languages, chat, instruction-following, regional tasks, and English capabilities.

View Leaderboard

Capabilities

Powering linguistic and cultural understanding across Southeast Asia

SEA-LION is designed to capture the intricate linguistic and cultural nuances of under-represented population groups and low resource languages across Southeast Asian communities.

Regional Excellence

  • Purpose-built on the latest State-of-the-Art open-model architecture and trained with Southeast Asian tokens to anchor regional representation on frontier design

Multilingual, Multicultural, Multimodal

  • Performant in 11+ SEA languages.
  • Supports image and text inputs, enabling visual Q&A, document comprehension, and culturally aware outputs

Agentic by Design

  • Built for agentic workflows (SEA-LION v4.5 onwards), with a custom speculative decoder delivering up to 6x greater inference efficiency

Resource Efficient

  • High performance even on low infrastructure
  • Quantised models are deployable on lightweight setups including laptops and edge devices

Safe & Secure

  • Safety-tuned for SEA contexts and complemented with safety guardrails – SEA-Guard
  • Aligned with human-curated data and calibrated to regional cultural norms and safety standards.

Open & Collaborative

  • Fully open-source and community-driven
  • Framework-agnostic for research and commercial use, available on Hugging Face, Ollama and all major hyperscalers

SEA-LION’s Guiding Principles

Choice

Users have the freedom to choose how to interact:

  • Closed vs Open
  • Build vs Buy
  • East vs West
  • Model
  • Independence

Open

Users have visibility on how it is being built:

  • Data
  • Models
  • Evaluations
  • Code Checkpoints

Responsible

Users understand the ethical reasoning behind how our models are:

  • Built
  • Aligned
  • Deployed

Inclusive

Users have a model that is relevant for their demographics:

  • Representative
  • Aligned
  • Collaborative