SINGAPORE, 4 AUG 2026 — Earlier today we reported that Vietnam has named large-scale Vietnamese language models a strategic technology product, putting them in the group reserved for things with an established market and tying tax incentives to the list. The region's most developed answer to that ambition already exists, and its details are more complicated than the word "sovereign" suggests.
SEA-LION is AI Singapore's family of open models for Southeast Asian languages, running since 2023 and supported by the Infocomm Media Development Authority. Its current release is the strongest evidence available anywhere in the region about what building for local languages buys you.
Where it wins, and by how much
Google, whose model SEA-LION v4 is built on, says it "ranks #4 out of 55 models on the SEA-HELM leaderboard, outperforming much larger systems while running on a laptop with 32GB RAM", and that it is the "#1 model for Tamil and Filipino".
Both halves of that matter, and they point in different directions. Fourth of fifty-five is a strong placing for a 27-billion-parameter model that fits on a workstation. First for Tamil and Filipino is the interesting part, because those are the languages where the general-purpose field is thinnest.
The margin is where the enthusiasm should stop. On Filipino, SEA-LION v4 is reported at 68.10 against 67.70 for Google's Gemma 3 27B — the model it was built from. Four tenths of a point. On Burmese, a variant built on Qwen leads at 49.56, and Burmese is a language where almost nothing else competes at all.
| SEA-LION v4 | Reading | |
|---|---|---|
| SEA-HELM overall | #4 of 55 | Strong for 27B parameters |
| Tamil | #1 | Under-served language, clear lead |
| Filipino | 68.10 vs 67.70 for base Gemma 3 27B | First place, by 0.40 |
| Burmese | 49.56, leading (Qwen-based variant) | Low absolute score; little competition |
Rankings and the #1 claims are Google DeepMind's, published about a model derived from its own Gemma. The Filipino and Burmese figures are as reported in secondary coverage of the SEA-HELM leaderboard; we could not read the leaderboard directly to confirm them. The right-hand column is RECATOOLS' reading, not any source's.
The pattern is consistent: regional tuning pays off for under-served languages, but its advantage shrinks to nothing for well-resourced ones. This is a useful result. It is also a much narrower claim than the one implied when a government names a national model as strategic infrastructure.
The benchmark does not cover everything the model claims
There is a gap between the languages SEA-LION says it handles and the languages it is ranked on, and it falls in an awkward place.
The model card lists ten Southeast Asian languages plus English: Burmese, Indonesian, Khmer, Lao, Malay, Mandarin, Tagalog, Tamil, Thai and Vietnamese. Google's description of SEA-HELM says the benchmark tests Burmese, Filipino, Indonesian, Malay, Tamil, Thai and Vietnamese — seven. Tagalog and Filipino are the same language for these purposes, so the three not measured are Khmer, Lao and Mandarin.
Mandarin is well served elsewhere and its absence from a Southeast Asian benchmark changes little. Khmer and Lao are the opposite case. They are among the lowest-resource languages in the set, the two with the least commercial reason for anyone to build for them, and therefore the two where a publicly funded regional model has the strongest justification for existing. They are also the two where the leaderboard cannot tell you whether it worked.
This is not a criticism of the model, which only claims support, not a score. It is a limit on the evidence. The "#4 of 55" placing covers seven languages; the case for regional models is strongest where the measurement stops.
What "sovereign" turns out to mean
Its provenance deserves more attention than it gets. SEA-LION v4 is built on Gemma 3, a Google model. Its model card places it under the Gemma Terms of Use — Google's licence, not an open-source one in the usual sense. Google's own account describes the training running on Google Cloud's Vertex Training Clusters. And the headline benchmark claims quoted above are published by Google, about a model derived from Google's, trained on Google's infrastructure.
None of that is concealed and none of it is improper. AI Singapore has done the work that matters for the region — continued pre-training across the languages, the evaluation harness, the low-resource coverage nobody else was going to fund. But a model with weights from an American company, under that company's licence, trained on that company's cloud, is not what the word "sovereign" brings to mind.
For Vietnam, which has just attached tax incentives to the category, that is the practical question its decision does not answer. A national language model built the way the region's best one was built inherits a licence and a supply chain from outside the region. Building one that does not means training from scratch, which is a different order of cost.
The disclosure on the model card
The most consequential sentences about SEA-LION v4 are not on any leaderboard. They are on its own model card, and they are unusually direct.
The card states that the model "was not tested for robustness against adversarial prompting", and that it "has not been aligned for safety". It notes the model can hallucinate and generate irrelevant content, and that its vision capabilities are comparable to the base Gemma 3 instruction-tuned model without improvement.
The developer is being straight about what it built, which is a credit to the project. A research release aimed at the open ecosystem is not a deployed consumer assistant, and saying so on the card is exactly right.
The risk is downstream. A model that has publicly been called the region's leading open LLM, that a minister can name in a speech, and that runs on a 32GB laptop, is going to be picked up by people who read the ranking and not the card. Any organisation putting it in front of the public owns the alignment and adversarial-robustness work itself, because the developer has said plainly that it did not do it.
What this says about the strategic-list approach
Vietnam's decision and Singapore's model are two answers to the same question, and together they describe the shape of the regional bet.
The case for it is now evidenced rather than assumed. Tamil, Filipino and Burmese are better served by a model built for them, something no commercial incentive was going to produce: the markets are too small and the training data too scarce. Public funding is the mechanism that works here, and it worked.
The case against is that the same evidence shows the advantage collapsing as the general models improve. Four tenths of a point on Filipino against the very model SEA-LION was derived from is not a moat. Whatever margin exists on the better-resourced languages is inside the range that the next general release can erase without anyone at Google or Alibaba thinking about Southeast Asia at all.
The defensible version of the strategy is narrow. It means funding the languages the market will not, admitting the advantage on others is thin and shrinking, and not confusing a derived model under a foreign licence with independence. Vietnam has put the ambition into an instrument with legal force. What it has not yet published is which of those two things it is buying.
View author profile → · Editorial policy
