Govt LLM – vision or illusion?

MALAYSIA has joined the ranks of countries building their own large language model (LLM) to play catch-up in the global race of artificial intelligence (AI).

Singapore and Thailand are also in the list.The Science, Technology and Innovation Ministry (Mosti) is leading the government’s initiative to develop a sovereign LLM tailored to the Malay language, which runs on local infrastructure, data and workforce.

In short, the LLM will be built and owned by Malaysians.

The idea looks dandy on paper but there is scepticism about the need for Malaysia to have its own LLM.

It will be costly especially if it involves localisation of the hardware or infrastructure.

Singapore is spending S$70mil for its larger LLM initiative while Thailand has allocated 120 million baht (US$3.38mil) for ThaiLLM. It is not known, however, whether the two countries are looking at lower costs with the recent breakthrough in LLM development by China’s DeepSeek.

Mosti Minister Chang Lih Kang says the government’s LLM will initially focus on the healthcare sector to use AI to enhance efficiency, improve patient outcomes, and support medical research.

He concurs that the LLMs require significant investment.

“For instance, a medium LLM, with parameters ranging between 10 billion and 100 billion, could cost up to US$20mil to develop,” he says in a reply to StarBiz 7,

To mitigate the costs, Mimos Bhd has partnered with Taiwan’s Phison to develop cost-effective technologies that will optimise the development process of the LLM.

Chang says that a medium LLM for the healthcare sector may take 18 to 36 months to be developed, depending on factors such as data availability, regulatory approvals, infrastructure, and expertise.

Costs are not the only concern. LLMs need a good pool of talent within the civil service to run and maintain.

There are also concerns over the use of personal data to train the model, considering that the government has access to extensive datasets.

RDS Partnership corporate lawyer Nur Shohidah Ramlee says there is a “genuine risk” that personal data could be used to make the LLM more compatible with the Malaysian user interface.

Taking these factors into account, does the country really need a government-owned LLM to integrate AI into public sector services?

Shouldn’t the job of developing an LLM be left in private hands, while the government focuses on regulating AI, ensuring ethics and fostering innovation?

Experts tell StarBiz 7 that they welcome the idea of a sovereign LLM to enhance public sector services, but opine that the government does not have to do it.

Public-private collaboration, where the government becomes the provider of data and the private sector develops the LLM, would be ideal, they say.

What is LLM?

In simple terms, an LLM is just like the human brain. Large volumes of data that are fed into it – a neural network – will be processed, allowing it to understand and generate human language.

This in turn allows the language model to power applications such as chatbots.

A government-owned LLM trained with public sector data and other local data seeks to reflect local languages and dialects, culture, laws and regulations.

Ultimately, this will help incorporate AI into public sector services in a meaningful way.

While many Malaysians know about OpenAI’s ChatGPT and China’s DeepSeek as well as Qwen, not many realise that Malaysia has three locally built LLMs.

YTL Power International Bhd, through YTL AI Labs, launched its LLM called Ilmu 0.1 last December. Ilmu 0.1 runs on Alibaba’s Qwen LLM. Based on YTL AI Labs’ website, Ilmu 0.1 has “successfully passed both Malaysian PT3 and SPM national exams in Bahasa Malaysia with As”.

Another listed firm, Agmo Holdings Bhd, has its LLM called Merdeka which is said to be Malaysia’s first crowdsourced AI language model. It offers AI Sovereignty as a Service (AI SaaS) to the government, corporate and enterprise sectors.

The third local LLM is by a startup known as Mesolitica, which has introduced its LLM called MaLLaM 2.5. It is based on Llama architecture while MaLLaM 2.5 Reasoning is based on Qwen architecture – both trained on Mesolitica’s datasets.

MaLLaM leverages Amazon Web Services’ (AWS) cloud infrastructure.

Using custom ML chips, including AWS Trainium, and AWS Inferentia, Mesolitica saw compute cost savings of 87% with a 5.5-fold increase in throughput (transactions per second) while training MaLLaM.

Of the three LLMs, Ilmu 0.1 is the standout based on the Malay MMLU benchmark. This is according to the data shared on the YTL AI Labs’ website.

In fact, Ilmu 0.1 has outperformed OpenAI’s GPT-4o and Meta’s Llama-3.1-70B. However, both GPT-4o and Llama-3.1-70B rank higher than Agmo’s Merdeka and Mesolitica’s MaLLaM.

MalayMMLU is the first multitask language understanding for the Malay Language.

The benchmark comprises 24,213 questions spanning both primary (year one to six) and secondary (form one to five) education levels in Malaysia, encompassing five broad topics that further divide into 22 subjects.The Madani government’s LLM is still in the early stages.

Mimos, an agency under Mosti, is working with strategic partners, including Mesolitica.

Speaking to StarBiz 7, Mesolitica chief executive officer Khalil Nooh says the startup is currently training data provided by Mimos.

“We are using local AI hardware acquired by Mimos. The LLM development for the government is still in trial stages and the government has not formally launched the initiative. How it will turn out depends on the scope and requirements of the government.

“We can also train MaLLaM using the government’s data instead of creating a new LLM for the purpose. But it is up to the government to decide,” he says.

Khalil did not disclose the cost of developing the LLM.

While experts think that an LLM for the civil service should be built through public-private partnership, with no or minimal capital expenditure burden on the government, there is the risk of official data leaks to the private sector.

A key question is, should private companies be given access to government data?

In response to this, Khalil notes that there are already standard operating procedures for government vendors to have access to higher tier data classifications.

Khalil believes that any concern on private sector access to government data can be addressed by having transparency on data set training from the beginning.

“Before feeding the data into the LLM, there should be a clear-cut agreement between both parties highlighting the do’s and dont’s.

“When there is clarity and regulation on how the data can be used, it removes the element of misuse,” he says.

Khalil also highlights the importance of having guardrails to filter certain elements from the LLM such as pornography, foreign propaganda and 3R issues of race, religion and royalty.

Prof Chan Chee Seng, the dean of Universiti Malaya’s Faculty of Computer Science and Information Technology, argues that the government must have at least one LLM of its own. However, it must be a sovereign AI model, he says.

Explaining further, Chan says a sovereign AI model will be built from scratch and not based on other LLM’s such as DeepSeek, Llama and Qwen.

“The government must own the hardware and software, only then can you entirely ensure the security of your data and eliminate reliance on a foreign party.

“That said, the government doesn’t necessarily have to build one as it can be a financial burden. This doesn’t include maintenance costs.”

Instead, Chan proposes a mechanism where a private sector player builds the LLM for the government, with the ownership in the government’s hands.

The LLM can then be licensed as a service to third parties such as agencies, businesses or foreign entities who want to utilise it, hence generating a revenue stream that can help pay for the cost of developing the model.

“For example, if a foreign company wants to do business in Malaysia and needs to use government services, the government can let them use the LLM for a fee.

“Once the upfront cost of the LLM developer is paid off, a revenue-sharing mechanism can kick in, generating revenue for the government over the longer term,” says Chan.

AI sovereignty

Chan leads YTL’s LLM, together with seven other Malaysian talents.

Interestingly, all seven are his former or existing students from Universiti Malaya. Four of them have yet to complete their undergraduate degree.

Chan says YTL’s Ilmu has yet to be sovereign, as it relies on Alibaba’s Qwen LLM.

“That is why it is referred to as Ilmu 0.1, instead of 1.0.

“We will be receiving Nvidia chips in the second quarter of this year and that would allow us to build our own LLM, without relying on external LLM.

“That will be sovereign. We are already working on it and we hope to launch Ilmu 1.0 in the second quarter,” he says.

Asked about the cost to build Ilmu 0.1, Chan says it is “low” as it uses an external foundational model.

“For now (Ilmu 0.1), we spend about US$10,000 to US$15,000 per month for the infrastructure, which is low. Part of the reason is, we used old graphics processing units (GPUs) or the A100 chips from Nvidia.

“Also, like DeepSeek, we have our own secret recipe to lower some of the costs.”

Once the sovereign LLM of YTL kicks in, Chan says the infrastructure costs will be higher due to the cost of more sophisticated GPUs.

“As for the operational costs including salaries, there won’t be a big difference. The cost of our sovereign LLM could be similar to DeepSeek or even slightly lower.”

The cost structure of YTL’s Ilmu could give some hints on how much the Madani government has to spend if it wants to develop its own LLM.

The more advanced the LLM is, the costlier it will be.

“There is also the cost for data as we will need to buy certain licences from Malaysian entities. How strong our LLM is will depend on the volume of data we feed into it,” he says.

According to Chan, the team is training 20 billion Malaysia data, but looks to progressively increase it to 50 billion as the more advanced Nvidia chips arrive.

For comparison, OpenAI’s GPT-3 was trained on a dataset that included approximately 300 billion tokens or words.

While OpenAI has not disclosed the number used to train GPT-4, the latest version of the model, it is reported to have been trained on trillions of tokens.

Technically, while local LLMs may not have the privilege of having large volumes of data like OpenAI, their advantage lies in being more accurate to local culture, language and other aspects.

This is because the LLMs are trained with specific Malaysian data that includes slang, which a foreign LLM like GPT 4 may not understand.

Chan foresees a future where there will be increased demand for local LLMs among companies doing business in Malaysia as they incorporate AI in their systems.

As more LLMs are created, including those owned by the government, he says there must be proper legislation to prevent misuse of data.

This includes a potential abuse by the government via censorship or propaganda.

“However, we cannot be too strict to the extent that it kills innovation,” Chan points out.

Meanwhile, RDS Partnership’s Nur Shohidah raises concerns about collaboration between Mosti and local companies in developing the LLM.

“Will these companies have access to sensitive personal data stored by the ministry?”

Another key legal concern on the development of LLM by Mosti is data confidentiality of data input by users.

Nur Shohidah says there is a risk that the data may be stored, processed, or even repurposed beyond the original user query.

“The model’s ability to retain contextual knowledge means that sensitive information could be integrated into future outputs, potentially exposing confidential or proprietary data.

“This raises significant concerns, particularly for industries dealing with confidential business communications, legal documents, or financial transactions. “If an LLM system trained on such data inadvertently discloses or replicates proprietary information, it could lead to breaches of confidentiality agreements and legal liabilities.

“Additionally, the concern is amplified as government agencies have access to extensive datasets, including individual records, company information, and national records.”

Nur Shohidah points out that while Mosti has committed to adhering to the National Guidelines on Artificial Intelligence Governance and Ethics in developing LLM, several legal concerns remain.

One of the key challenges is that the guidelines lack the force of law, unlike legislated statutes, they do not carry direct penalties for non-compliance.

This raises concerns about enforcement, particularly in ensuring transparency and confidentiality in the development and deployment of LLM.

Furthermore, Malaysia lacks an independent regulatory body with investigatory or enforcement powers to oversee LLM effectively.

“Without a comprehensive legal framework to address these issues, the reliance on ethical guidelines alone may be insufficient to safeguard against potential risks and ensure responsible LLM development,” she argues.