Advancing Southeast Asia’s AI Future Through Sovereign AI Models

In late December 2025, Grok – the AI model developed by Elon Musk’s company xAI – released a feature that allowed users to generate non-consensual sexualised images of women and children. Within a few weeks, in a unique move, Indonesia, Malaysia and the Philippines banned Grok after they deemed xAI’s initial mitigations to be insufficient. The countries only removed the ban when xAI placed new restrictions on Grok’s image generation capabilities. While this episode was resolved, it underscores the risk that Southeast Asian countries face when depending on foreign AI models, particularly when the US is currently limiting the regulation of Big Tech.

In the wake of this maelstrom, the opportunely timed AI for Developing Countries Forum gathered in early February 2026 to discuss AI sovereignty. At its Bangkok summit, delegates from over 100 countries adopted a declaration to pursue AI sovereignty, which entails having the ability to control AI data, computing resources and, crucially, the models operating within their borders. A critical component of AI sovereignty involves developing and retaining local talent so that countries will not simply be consumers but rather creators of AI models and applications. And, as articulated in the Bangkok Declaration, owning AI model creation is necessary for ensuring that models reflect a nation’s “languages, cultures, and values”. More specifically, most countries seek AI sovereignty through control of data and models fundamentally because the choice of data used and how the models are trained dictate the AI outputs consumed by their citizens.

In alignment with global trends across Europe, the Middle East, and the Asia-Pacific, several Southeast Asian nations have taken steps towards developing sovereign AI models. In December 2023, Singapore identified the “strategic need to develop sovereign capabilities in LLMs” and launched its SEA-LION models. Malaysia launched its domestically developed ILMU LLM in August 2025, and Vietnam passed its AI Law in December 2025, emphasising sovereignty over AI data, infrastructure, and models.

The cost-effective way to develop a domestic LLM is to build it with an open-source “foundational” LLM like Alibaba’s Qwen or Meta’s Llama. Leveraging open source models requires fewer data, computing, and personnel resources; this approach also provides access to the latest functionality and security features. SEA-LION takes this approach: it “pre-trains” existing foundational LLMs using data from 13 regional languages, then “fine-tunes” them in a regionally appropriate manner. Recent tests show that SEA-LION performs well in Southeast Asian language performance tests compared to major LLMs. SEA-LION’s regionally trained LLMs, which are also open source, have been subsequently used as base models for other countries’ LLMs, such as Indonesia’s Sahabat-AI.

The other path towards AI model sovereignty requires a country to invest in building an LLM from scratch, starting at the foundational level. In Southeast Asia, only Malaysia has taken this approach. Its ILMU model is developed by YTL Corporation, an infrastructure and technology conglomerate working in collaboration with Universiti Malaya. While ILMU requires considerably more resources to develop, it provides the most sovereignty, being insulated from future disruptions if open source foundational LLMs change their terms. Furthermore, some “open source”-branded LLMs like Llama are not truly open source because they do not release their full source data and codebase. This precludes their interpretability and limits the level of assurance ideally required for critical sector applications.

Sovereign AI models do not need to outcompete the major LLMs in all dimensions.  Rather, Southeast Asian nations can carve out a segment for locally relevant, domestic LLMs by leaning into their inherent regional strengths.

Irrespective of the approach, the key risk for sovereign AI models is whether they can develop the needed performance and functionality that will then drive sufficient adoption that justifies the significant investments required. Major LLMs such as ChatGPT have early mover advantages and dominate market share. This challenge is amplified in smaller countries that have fewer resources and smaller customer bases.

Fortunately, sovereign AI models do not need to outcompete the major LLMs in all dimensions.  Rather, Southeast Asian nations can carve out a segment for locally relevant, domestic LLMs by leaning into their inherent regional strengths.

To begin with, regional model builders are closest to regional customers, and they can and need to relentlessly “get out of the laboratory” to understand and solve local customer problems. Success is not only measured by high performance scores, but ultimately by adoption. Malaysia’s ILMU is a good example when it launched with established partners in e-commerce, media, and telecommunications. Indonesia’s Sahabat-AI is also strategically partnering with the government – a future anchor customer – to develop a citizen services platform. Taking advantage of the fact that AI is a nascent technology, Singapore launched its 100 Experiments programme, which provides funding to uncover new customer use cases.

Next, quality data is the currency for success. Domestic LLMs need to make the acquisition of local data – a clear geographic advantage – part of their core strategy. Malaysia’s ILMU employs a large national team to acquire data from domestically licensed and Malaysia-centric sources such as educational and government materials. In this process, ILMU is able to ensure that all training data resides on the sovereign AI data infrastructure within domestic borders.

Finally, countries should lean into ASEAN’s cooperative mandate – existing AI plans already emphasise interoperability, and consistency in standards will make it easier for regional LLMs to scale. With cooperation, smaller countries too can deploy sovereign AI models by adapting regionally contextualised open source LLMs. Countries lacking the workforce or resources to retrain a foundational model could enter into regional data governance agreements for collaborative training. A regional approach, though without formalised inter-state governance, is currently employed to train Latam-GPT, an LLM jointly developed by 60 participating institutions across 15 countries in Latin America. Malaysia’s stated intention to be a regional AI hub – along with its investments in data and computing infrastructure – provides the building blocks for such an approach.

Despite the risks and hurdles of deploying sovereign AI, the rewards can be attained when countries lean into their strengths and uniqueness. In doing so, they will position themselves to participate in the full potential of this transformative technology while serving the specific needs of their people with respect to their diverse languages, cultures, and values.

2026/53