KUALA LUMPUR (June 25): YTL AI Labs launched its proprietary large language model (LLM), ILMUchat, marking a milestone in Malaysia’s sovereign artificial intelligence (AI) landscape.

While global technology conglomerates dominate the generative AI sector, YTL AI Labs is aiming to capture the domestic enterprise and consumer markets by addressing a technical gap seen involving the nuance of local language, mixed dialects and cultural context.

A core architectural focus of ILMUchat is its ability to natively process and generate text in the highly colloquial, multi-lingual environment characteristic of daily Malaysian communication.

YTL AI Labs CEO Foong Chee Mun said unlike mainstream global models that lean heavily on formal language structures, ILMUchat has been specifically optimised to handle rojak speech patterns, where elements of Bahasa Melayu, English, Chinese, local dialects and Manglish are naturally blended within a single conversation.

He added that global LLMs inherently struggle with linguistic code-switching in minor or regional language markets.

"When you talk in minor languages, global models tend to deviate from it," said Foong.

"If you talk to it in English and Malay, sometimes it will switch to speak to you in Korean for no good reason. That is something inherent to most large language models. We have trained ILMUchat to follow the local language flow precisely."

The platform's features extend beyond basic text generation. The pro version allows users to build mini-applications and design presentation slides directly inside the ecosystem.

To build this capability, YTL AI Labs engineering team fed the model vast amounts of localised data and chat applications. This specific data ingestion allowed the model to interpret highly nuanced local abbreviations, from institutional acronyms like PTPTN and KWSP, to differentiating between kuih lapis made in Kuala Lumpur and Sarawak.

ILMUchat currently has over 10,000 organic users and internal tracking data indicates that initial users are primarily utilising the model for translation purposes, between Malay and English, to understand multiple documents and to create a summary out of it.

Taking this into account, Foong said this is also why it is important for the model to understand the contextualised content.

“Beyond that, a lot of users are using it to generate things, like posters for their businesses or events because it is very good at generating localised content,” he said.

“For example, if someone tells the model ‘Hey, there is a neighbourhood gathering this weekend and we will serve nasi lemak and kuih lapis’, they will be presented with nasi lemak and kuih lapis images specific to Malaysian context.”

YTL AI Labs is positioning ILMUchat as a critical infrastructure asset for data sovereignty. For local enterprises and government agencies, relying entirely on foreign cloud-hosted AI application programming interfaces (APIs) introduces legal and operational risks.

An example, Foong highlights, is the extraterritorial reach of foreign legislation, such as the US Cloud Act (Clarifying Lawful Overseas Use of Data Act). The federal law grants US authorities the power to compel US-based technology companies and service providers to produce stored data, regardless of whether that data is physically stored in the United States or abroad.

"Under the US Cloud Act, even if a foreign provider's cloud servers are physically located in Malaysia, the data remains subject to US jurisdiction, meaning their government can legally request access to that information," Foong said.

“That directly jeopardises domestic data sovereignty."

On top of that, operational availability remains an ongoing geopolitical risk. If a foreign state or provider decides to restrict API access during a trade dispute or regulatory shift, domestic organisations relying on those models could face sudden operational paralysis. Foong says a sovereign model ensures continuous local uptime under Malaysian jurisdiction.

The engineering team’s focus was to achieve a hyper-local identity without sacrificing global baseline intelligence, which required a layered training strategy.

Foong said they designed the underlying ILMU models to remain competitive on a global scale by targeting top-10 performance rankings in open-source benchmarks for non-localised disciplines such as physics, biology and advanced coding.

To layer the Malaysian identity, the development team utilised several training pipelines, such as supervised fine-tuning, which trains the model on curated question-and-answer pairs designed to mirror Malaysian conversational structures and reinforcement learning, which implements a strict reward-and-punishment feedback loop to bring localised knowledge to the forefront of the model's retrieval priorities.

"We don’t know every single permutation a user is going to ask, so we train it structurally to remember its identity: 'Remember you are Malaysian, speak like a Malaysian’,” Foong explained.

“It is a profound engineering challenge. Every single training run takes roughly 20 to 25 days across hundreds or thousands of dedicated graphics processing units (GPU)."

Beyond vocabulary, localising an LLM requires aligning its safety parameters with national sensitivities. YTL AI Labs spent more than six months developing custom guardrails to govern ILMUchat’s outputs, particularly concerning topics that could disrupt multi-ethnic stability.

"Malaysia is inherently a multi-ethnic country, and our societal boundaries are very different from the United States or China," Foong said.

"We built specific guardrails for Malaysian sensibilities so that if a user inputs an unsafe or inappropriate prompt, the system safely restricts the output based on compliance parameters that we as a nation collectively value."

While the company hopes to eventually scale the technology to benefit all 35 million Malaysians via its free basic tier, the immediate distribution engine relies heavily on Yes’ infrastructure.

Through a strategic partnership with YTL Communications’ mobile brand, Yes, the platform’s pro version is being deployed to over three million users nationwide, establishing Yes as the country's first "AI-first telco".

Under the commercial agreement, all active Yes subscribers will receive complimentary access to the ILMUchat Pro plan, which retails to the public at RM50 per month, integrated directly into the MyYes App.