While the biggest AI labs race to build ever larger models in English and Mandarin, a small Hong Kong startup is making the case that one of the world's most spoken languages has been left behind. Votee AI, led by founder and chief executive Pak Sun Ting, builds large language models trained specifically for Cantonese, a language used by more than 80 million people that has been treated as a rounding error by most of the industry's biggest players.
The company's approach is practical rather than building foundation models from scratch. Votee AI takes open source models from Meta and Alibaba and retrains them on Cantonese specific data, producing models with as many as 70 billion parameters tuned for the language's grammar, slang, and cultural context. That retraining strategy lets a small team compete on capability without needing the enormous compute budgets that frontier labs spend building models from the ground up.
Where the data comes from
Feeding a Cantonese model requires a very different data pipeline than an English or Mandarin one, since far less Cantonese text exists online relative to those languages. Votee AI has grown its Cantonese corpus from around 100 million tokens to more than 500 million, pulling from sources like broadcaster Radio Television Hong Kong, contributions from universities and community groups, and synthetic Cantonese data the company generates in house to fill the gaps. Ting has framed the stakes plainly, arguing that a language left out of AI systems risks being left out of the tools people increasingly rely on for everyday tasks.
Cheap by design
Perhaps the most striking number is cost. Votee AI says it can train one of its Cantonese models for around 250,000 dollars, a fraction of what frontier labs spend building general purpose models. Benchmarks reportedly show the models handling everyday Cantonese conversation reasonably well, while still stumbling on more localized cultural knowledge, a reminder that language coverage and cultural fluency are not quite the same problem.
A profitable niche, not just a passion project
Votee AI is not running on goodwill alone. The company says it is already profitable, with banks, universities, and government departments among its clients, an unusually early revenue base for an AI startup this specialized. Hong Kong businessman Allan Zeman advises the company, adding a well known local name to a startup that is explicitly positioning itself around Hong Kong identity and language rather than trying to out build global AI giants at their own game.
Part of a bigger sovereign AI trend
Votee AI's pitch fits into a broader movement sometimes called sovereign AI, the idea that regions and language communities should control the AI systems trained on their own data and culture rather than depending entirely on models built elsewhere for someone else's market. For Cantonese speakers, that has meant watching English and Mandarin dominate AI development while their own language gets treated as a niche afterthought, if it gets addressed at all.
Eyeing Southeast Asia next
Votee AI is now looking beyond Hong Kong, with active discussions underway with AI Singapore as the company eyes expansion into Southeast Asian markets. The region is home to dozens of languages that face the same imbalance Cantonese does, spoken by tens of millions of people yet underserved by AI models built around English, Mandarin, or a handful of other dominant languages. If Votee AI's model of retraining open source systems on underserved languages works commercially in Hong Kong, it offers a template other language communities across Southeast Asia could plausibly follow.
