SEA-LION
Multilingual language models for Southeast Asia
SEA-LION is a family of large language models built by AI Singapore to support languages in the Southeast Asian region. Released in 2025, these open-weight models address the under-representation of low-resource languages in artificial intelligence.
- Supports 11 Southeast Asian languages
- Released under an open MIT licence
- Based on Gemma2 and Llama architectures
- Context length of 8192 tokens
- Developed by AI Singapore
What it is
SEA-LION stands for Southeast Asian Languages In One Network. It is a collection of large language models developed by the Products Pillar at AI Singapore. The project aims to serve the Southeast Asia region by supporting eleven local languages alongside English and Chinese. The family includes variants based on Gemma and Llama architectures, such as the 9-billion parameter Gemma-SEA-LION-v3-9B-IT. These models are designed for general language tasks and instruction following within their specific linguistic domain.
How it came about
The research team introduced the SEA-LION family to address the gap in representation for low-resource languages. Most prior development focused heavily on English. The group released their work in April 2025, with revisions continuing through October of that year. The paper describing the models was accepted at the IJCNLP-AACL 2025 conference. The team open-sourced the weights to benefit the wider community in the region.
Inside the model
The models use a decoder-only transformer design. The Gemma variant employs the Gemma2 architecture and its default tokenizer. Training involved large-scale multilingual continued pre-training. This was followed by a post-training regime with multiple stages of instruction fine-tuning, alignment, and model merging. The instruction tuning covered English and ASEAN languages like Indonesian, Thai, and Vietnamese. The context length for the Gemma variant is 8192 tokens.
What it can do
The models support Burmese, Chinese, English, Filipino, Indonesian, Javanese, Khmer, Lao, Malay, Sundanese, Tamil, Thai, and Vietnamese. Evaluations used the SEA-HELM benchmark for tasks such as question answering, sentiment analysis, and translation. The instruct models were also tested on localized versions of IFEval and MT-Bench. Results indicated state-of-the-art performance among models supporting these languages. The models do not support system prompts or tool calling.
How to use it
AI Singapore released the model weights under an MIT licence. This permits commercial use and modification. Users can download the checkpoints from Hugging Face. The organisation encourages researchers and developers to contribute data, report bugs, or train new versions. Older revisions of the models are accessible via revision parameters on the hosting platform.
Safety and testing
The released models have not been aligned for safety. The developers state that users must perform their own safety fine-tuning and security measures. The team warns that the models can hallucinate or generate fictional content. They advise caution when interpreting responses due to potential reasoning inconsistencies. The authors accept no liability for damages arising from the use of the weights.
Questions people ask about SEA-LION
Which languages does SEA-LION support?
The model supports eleven Southeast Asian languages: Burmese, Chinese, English, Filipino, Indonesian, Javanese, Khmer, Lao, Malay, Sundanese, Tamil, Thai, and Vietnamese.
Is SEA-LION safe to use without modification?
No. The developers state that the models have not been aligned for safety. Users should perform their own safety fine-tuning and security measures before deployment.
What is the licence for SEA-LION?
The models are released under the MIT licence. This is a permissive open-source licence that allows for commercial use and modification of the weights and code.
Good reading and sources
There is no English Wikipedia article on SEA-LION yet, so the facts on this page come from the sources listed above: huggingface.co/aisingapore/Gemma-SEA-LION-v3-9B-IT, arxiv.org/abs/2504.05747. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.