SEA-LION

SEA-LION is an open-source LLM family, not a single model. It includes multiple sizes (3B, 7B/8B, 70B), multiple base architectures (originally in-house, then based on Llama 3 and Gemma from v3 onwards via continued training), and multiple variants (base, instruct fine-tuned, RAG-adapted).

On the technical stack:

• Training data: centred on the 11 official Southeast Asian languages (English, Chinese, Malay, Indonesian, Thai, Vietnamese, Filipino, Tamil, Burmese, Khmer, Lao); training corpus around 1 trillion tokens, with SEA languages far over-represented compared to general LLMs
• Base model choice: v1 self-built architecture → v2 based on Llama 2 → v3 based on Llama 3 / Gemma with continued pre-training and instruction tuning
• Compute: relies on the Singapore National Supercomputing Centre (NSCC) and sponsored compute from Google Cloud / AWS
• Open-source licence: MIT / Apache, commercially friendly, allowing direct enterprise use
• Companion tools: SEA-HELM (evaluation benchmark) and SEA-Guard (safety) form the complete tooling chain

Models can be downloaded directly from HuggingFace, or accessed via the official sea-lion.ai API. It is one of the few LLMs that is produced by a national-level institution yet fully open and explicitly designed to encourage commercial use.

SEA-LION holds a very clear position in the LLM ecosystem: "the SOTA foundation for Southeast Asian languages".

The core technical problem it solves: general LLMs collapse on smaller Southeast Asian languages. GPT-4 might score 95 on English/Chinese tasks but drops to 30–40 on Burmese, Khmer, or Lao (reproducible on SEA-HELM). The root cause is training data — SEA languages typically make up under 1% of general LLM training corpora.

SEA-LION's approach is continued pre-training:

• Take a strong base model with general capabilities (Llama 3 / Gemma)
• Continue pre-training with large amounts of SEA language corpora to restore semantic fidelity in smaller languages
• Without sacrificing too much English capability (the technical challenge)

Once this works, SEA-LION beats same-sized Llama 3, Gemma, and Qwen on Southeast Asian language tasks in SEA-HELM — its most compelling hard evidence.

At a broader level, SEA-LION is also an important case study for "regional adaptation of open LLMs". It proves: not every country needs to train its own GPT-4, but every language region may need its own continued pre-training variant — a pattern Indonesia, Malaysia, and Vietnam are now imitating.

SEA-LION is the most symbolically important output of Singapore's AI strategy — clearer than any policy document on "what kind of AI Singapore wants to do".

In the seven-lever framework, SEA-LION sits across three levers:

• Lever 5 (government adoption): government agencies deploy localised AI services on SEA-LION, avoiding sending data to overseas big tech
• Lever 6 (international): SEA-LION is Singapore's "tech calling card" at ASEAN AI cooperation, GPAI, Bletchley/Seoul summits — proof that small countries can produce globally usable open-source models
• Lever 3 (industry adoption): once open-sourced, local enterprises (especially in finance, government, healthcare with sensitive data) can fine-tune directly without depending on overseas APIs

A take: SEA-LION's real value is not in its benchmark numbers but in being a "sovereign AI" reference project — it tells Southeast Asia: "you can also have your own LLM foundation, you don't have to use only OpenAI". This narrative value far exceeds its lift on any single benchmark.

But SEA-LION has real bottlenecks too:

• Not trained from scratch — it depends on Meta/Google open-source bases (Llama 3 / Gemma); if those move closed-source, the whole project must restart
• Resources far smaller than big tech — AISG's compute budget is roughly 1/100 of big tech, iteration speed is naturally slower
• Commercial loop unproven — currently mostly used by government and open-source community; enterprise paid scenarios have not scaled

These bottlenecks are exactly the questions to be answered in the NAIS 2.0 era — should SEA-LION v4/v5 move to a self-built base? Should compute be bound to a regional GPU cluster?