Cover Khalil Nooh, co-founder and CEO of Mesolitica is one of the pioneers behind MaLLaM (Photo: Mesolitica)
Khalil Nooh, co-founder and CEO of Mesolitica is one of the pioneers behind MaLLaM (Photo: Mesolitica)
1 / 1
At the intersection of innovation and identity, Khalil Nooh’s MaLLaM brings Bahasa Melayu into the AI age; paving the way for a more inclusive digital future
As the global race to innovate reaches a fever pitch, Khalil Nooh is steadily working towards democratising AI for the Malay-speaking world. The co-founder and CEO of Mesolitica speaks with the quiet confidence of someone who has spent the better part of seven years building something that didn’t exist before: a large language model that understands the nuances of Malaysian culture, including local dialects, slang, colloquialisms and 16 regional languages. To Khalil, localisation was a necessity to ensure Malaysia would not be left behind in the AI age.
“Even before AI, when it came to programming, our local folks still had to work on understanding English on their own to access quality learning content from Ivy League institutions,” Khalil explains. “Only if you already had good English skills were you able to have access to this knowledge. But for non-English speakers, it’s always been a game of waiting for a reliable translator.”
Read more: AI decoded: what those buzzwords really mean
This digital divide became one of the main catalysts for Mesolitica to pioneer MaLLaM (a B2B Malaysian Large Language Model), which launched in December 2023 as the first large language model trained from the ground up on nearly 200 billion tokens of Malay-specific content from at least 197 databases, over 349GB in size. The journey began in January 2019, after the Cambridge Analytica case study demonstrated the powerful capabilities of social media analytics in political contexts. The firm’s use of Facebook data to create detailed user profiles for political micro-targeting showcased the sophisticated potential of large-scale data analysis.
Recognising these technological possibilities, Khalil’s co-founder and CTO, Husein Zolkipli, saw an opportunity to develop similar (and more ethical) analytical capabilities, specifically focused on the Bahasa Melayu language. “When I joined Mesolitica, it was to continue using AI to see how we could build useful applications around local language,” he says. “When ChatGPT arrived on the scene, we had been building our own chatbot. But when ChatGPT reached 100 million users, we saw the need to pivot into LLMs.”
Above MaLLaM: 128k context length, multi-lingual including malay dialects chat language model, function call, Malaysian centric and OpenAI compatible
MaLLaM: 128k context length, multi-lingual including malay dialects chat language model, function call, Malaysian centric and OpenAI compatible
1 / 1
MaLLaM began with Microsoft and OpenAI GPT-4 APIs, but the breakthrough came when Meta released Llama 2 in mid-2023. “That was the starting point for us. We saw that we could now create our own LLMs because of this open-source model from Meta,” Khalil says. What followed was an ambitious eight-month local data collection effort, crawling Malaysian websites and preparing datasets.
The actual training of MaLLaM 1.0 took just 10 days, powered by 80 Nvidia A100 GPUs on Microsoft Azure over the Christmas period. “We had access to these Nvidia GPUs through our participation in the Microsoft startup programme,” he notes, emphasising the crucial role of strategic partnerships in Malaysia’s AI ambitions.
See also: AI or Human? GPT-4.5 proves it’s hard to tell the difference
The data collection process itself reads like a digital archaeology project. Following the methodology of tech giants like Meta and OpenAI, Mesolitica focused on Malaysian domains: government websites, local forums and social media content. “We do open source our datasets on GitHub, so if anyone wants to remove their content, they can,” Khalil admits, acknowledging the ongoing legal complexities surrounding fair use in AI training.
Perhaps most intriguingly, Mesolitica has forged partnerships with local institutions to access uniquely Malaysian datasets. Their memorandum of understanding with RTM radio provides access to state-specific dialect recordings with a view to build an ElevenLabs-equivalent platform for a range of possible use cases like voice cloning for digital newscasters. Meanwhile collaborations with local call centre operators offer telephone-quality conversations perfect for training AI voice agents. “This is especially great for dialect training,” Khalil explains, addressing a pain point familiar to any journalist who has struggled with transcription software that cannot recognise Malaysian accents.
Beyond commercial success, Mesolitica is also exploring training. Through partnerships with government agencies like the Department of Skills Development (JPK), Mesolitica is conducting technical training sessions, teaching Malaysians to build their own AI applications. “It’s capacity building,” Khalil notes, though he initially resisted the idea of teaching potential competitors. “We are teaching people to build the things that we are building.”
Khalil’s vision extends far beyond mere technological prowess. “The future is solo. In the AI era, it’s possible for one person to build a billion-dollar company.” This isn’t mere Silicon Valley hyperbole: Khalil points to Base44, a six-month-old solo-founder company recently acquired by Wix for USD $80 million.
The key, he argues, lies in combining “agentic AI”, that is AI that autonomously makes decisions and acts to achieve goals without human guidance, with “vibe coding”, where a developer can instruct an AI in plain language to write and fix code. Khalil sees this as a democratisation of programming that allows non-coders to build applications. “Non-coders actually have an unfair advantage. They bring in their own domain knowledge and insights that the programmer won’t know,” he says. “Once you do get into vibe-coding, you build confidence to create. One brick at a time.”
Don’t miss: AI’s creative conundrum: From copyright infringement to artistic originality
As Malaysia’s AI landscape evolves, Khalil has made digital independence a cornerstone of his strategy. When Amazon Web Services (AWS) launched its Kuala Lumpur region, Mesolitica migrated everything from Jakarta and Singapore to run locally. “We needed to be able to say that everything that we do in Mesolitica runs 100 per cent in Malaysia,” he explains. “That is very important to us if we are championing the AI sovereignty agenda.”
Looking ahead, Mesolitica’s roadmap is ambitious. With access to new H100 GPU clusters in Malaysia, the team has already trained MaLLaM to 70 billion parameters—a significant leap in model capability. The immediate focus is on launching their speech-to-speech offering, creating AI voice agents that can handle real-time conversations in Malaysian dialects. Their partnership with a local client aims to deploy 400 AI voice agents for customer support by year’s end, starting with simple tasks like appointment reminders before expanding to more complex interactions.
“We’re taking it slow,” Khalil explains, mindful of the technical challenges involved in creating voice agents that can understand Manglish, operate in real-time with millisecond response times, and meet the stringent requirements of regulated industries like banking.
For young Malaysians aspiring to enter the AI space, Khalil’s advice is refreshingly direct: “Start building. First, be a power user of ChatGPT and then start building. If you’re not using ChatGPT every day, then you won’t know what the limitations are and what’s possible.”
His own workflow exemplifies this philosophy, though he’s mindful of AI’s potential cognitive trade-offs. “When we started to use mobile phones, we forgot all our phone numbers. It’s the same thing,” he observes, referencing recent MIT research on AI’s impact on memory, attention and problem solving. His solution? “Instead of always prompting to get an answer, you need to start prompting so that the responses that you get will encourage you to do follow-ups—that means you continue to become curious.” Rather than seeking instant solutions, he advocates using AI as a “companion to lead you down the rabbit hole of curiosity-driven exploration.”
Topics
Best of Tatler Asia video highlights
Featured videos from around Tatler Asia: Get exclusive behind-the-scenes look at the interviews we do, the events we attend, the shoots we produce, and the incredibly important people who are part of our community
