Singaporean AI Models Built for Southeast Asia

Most popular language models learn from mountains of English and Chinese text, so they tend to fumble when a chat switches between Malay and English mid-sentence, or when someone writes in Singlish. Singaporean research teams have spent the last few years closing that gap, and the result is a small but capable group of open models tuned for the region's languages, accents and cultural habits. Here is a quick look at the main ones before we go through each in detail.

SEA-LION

SEA-LION is the flagship of the national effort, led by AI Singapore and open source from the start. Rather than training from scratch, the team continues pre-training strong open models and adapts them to the region. Earlier releases built on Gemma 2 and Llama 3.1, while the v4 generation moved to Qwen3 and Gemma foundations. Newer mixture-of-experts variants have also appeared, including a larger one built with NVIDIA.

Features

• Covers more than ten regional languages, including Indonesian, Malay, Thai, Vietnamese, Tamil, Javanese and Sundanese, alongside English and Chinese
• Offers a 128,000-token context window in recent versions, so long documents stay manageable
• Switched to a byte-pair encoding tokenizer in v4 for more efficient handling of regional scripts
• Includes reasoning-capable and instruction-tuned versions, with sizes small enough for a laptop with 32GB of RAM

Strengths

• Handles code-switching and informal chat far better than general-purpose models
• Works well as a companion to larger models, filling in regional context they miss
• Open weights and a commercial-friendly approach lower the cost of building local products

SEA-LION Guard

This is the safety sibling of SEA-LION. Instead of writing text, it acts as a decision model that returns probabilities for categories such as safe, sensitive and unsafe, and it also supports rule checks for AI agents. The v2 versions are fine-tuned from compact Gemma models.

Features

• Outputs structured probabilities rather than free text
• Available in small sizes that suit edge and low-cost deployments
• Designed around Southeast Asian sensitivities rather than Western defaults

Strengths

• Fast enough to sit in front of a chatbot without noticeable lag
• Catches region-specific risks that generic filters overlook

MERaLiON

MERaLiON, short for Multimodal Empathetic Reasoning and Learning in One Network, comes from A*STAR's Institute for Infocomm Research. Its first release, MERaLiON-AudioLLM, fuses a Whisper-based speech encoder fine-tuned on local data with SEA-LION v3, giving a model of roughly 10 billion parameters that takes audio and text together. It was trained partly on Singapore's National Speech Corpus.

Features

• Accepts audio and text input and produces text output
• Version 2 added Malay, Tamil, Bahasa Indonesia, Thai and Vietnamese on top of English, Mandarin and Singlish
• Picks up on emotional cues in pitch, tone and volume
• A separate speech encoder was trained on 200,000 hours of audio, including 10,000 hours of Singapore speech

Strengths

• Understands local accents that mainstream speech systems often mishear
• Suited to customer service, elder-care check-in calls and scam-call detection, where tone matters as much as words
• Backed by a consortium that includes banks, telcos and government agencies, so real deployments are already being tested

Sailor2

Sailor2 comes from Sea AI Lab and a wider community of academic and industry partners. It is a family of 1B, 8B and 20B models expanded from Qwen2.5 and further pre-trained on about 500 billion tokens. The authors report that the 20B model reaches a 50-50 win rate against GPT-4o on Southeast Asian language tasks.

Features

• Supports 13 regional languages while keeping solid English and Chinese ability
• Released under the Apache 2.0 licence
• Ships with a full cookbook, data-cleaning scripts, an evaluation suite and the SEA-WildBench chat benchmark

Strengths

• Strong coverage of lower-resource languages such as Khmer, Lao, Burmese and Cebuano
• The three sizes make it easy to match the model to the hardware
• The open training recipe helps other teams build models for their own underserved languages

LionGuard 2

GovTech Singapore's responsible AI team built LionGuard to deal with a specific problem: moderation tools trained on Western data miss locally harmful phrasing. The original version targeted Singlish using a dataset of about 138,000 labelled texts. LionGuard 2 expanded to English, Chinese, Malay and partial Tamil.

Features

• Uses pre-trained embeddings with a multi-head ordinal classifier instead of a heavy fine-tuned LLM
• Runs on CPUs and can be retrained in under two minutes
• Weights and part of the training data are public

Strengths

• Outperformed commercial and open-source moderation systems across 17 benchmarks
• Already deployed on the government's AI Guardian platform
• Cheap to run, which makes it practical for small teams adding a safety layer

Choosing between them

If the job is general text work in regional languages, SEA-LION or Sailor2 is the natural starting point. For voice products, MERaLiON is the one built for the task. And any public-facing chatbot in the region benefits from a local guardrail like LionGuard 2 or SEA-LION Guard sitting in front of it. Because these projects update often, it is worth checking each model's repository for the latest versions before building on them.

› More Article