Hong Kong startup Votee AI builds Cantonese model to challenge English-Mandarin dominance

Votee AI, headquartered in Hong Kong, is developing large language models that operate in Cantonese, a language spoken by over 80 million people but largely ignored by mainstream AI. The firm adapts open-source models such as Meta’s Llama and Alibaba’s Qwen, adding Cantonese corpora scraped from Radio Television Hong Kong, academic partners and synthetic datasets, expanding its token base from 100 million to more than 500 million.

Its 70-billion-parameter model, trained on roughly $250,000 worth of compute, is being sold to banks, universities and government departments for tasks in education, healthcare and policing. Ting positions the effort within a broader “sovereign AI” movement, noting that smaller, locally-focused models can meet many public-sector needs without relying on massive English or Mandarin systems. The startup is profitable, counts tycoon Allan Zeman as an advisor, and is in talks with AI Singapore to broaden its reach across Southeast Asia. Ting also envisions applying the technology to preserve endangered languages worldwide.