Without local languages, ‘AI is essentially useless’: Hong Kong’s Votee AI is taking on English and Mandarin’s AI dominance with a Cantonese model

The rise of artificial intelligence (AI) is predominantly shaped by two languages: English and Mandarin Chinese. This linguistic dominance is evident in the leading AI developers, such as OpenAI, Anthropic, DeepSeek, and Z.ai, who are primarily located in the United States and mainland China. Consequently, their AI models are optimized for performance in these native languages, leaving a significant number of other languages and dialects—some spoken by tens of millions of people—neglected in the process.

Pak-Sun Ting, the CEO of Votee AI, a startup based in Hong Kong, highlights this disparity, stating, "The whole AI revolution is in English and Mandarin. There’s only a very small fraction that represents other languages." Votee AI is among a growing cohort of companies striving to address the oversight of AI in relation to languages beyond the dominant two. The company focuses on developing AI models specifically for Cantonese and other underrepresented languages.

The Challenge of Cantonese

Votee AI employs open-weight models from established developers such as Meta and Alibaba. These models are retrained using Cantonese data, which the company then markets to various sectors, including banks, universities, and government departments. Ting emphasizes the importance of incorporating Cantonese into AI applications, stating, "If those don’t get covered, then AI is essentially useless."

Cantonese, often categorized as a dialect of Chinese, is fundamentally distinct from Mandarin. The grammatical structure and vocabulary of Cantonese differ significantly, especially in Hong Kong, where speakers frequently blend English and Cantonese within the same conversation. Despite being spoken by over 80 million individuals—comparable to the number of Korean speakers and exceeding those who speak Italian or Thai—the availability of standardized written data for colloquial Cantonese remains limited.

Research indicates that mainstream AI models are not entirely ineffective at processing Cantonese. The HKCanto-Eval benchmarks, developed by a collaboration between Kyushu University, the Education University of Hong Kong, and the local AI community hon9kon9ize, reveal that while these models can manage everyday Cantonese to a reasonable extent, they often struggle with cultural nuances and local knowledge.

Ting describes the process of creating a Cantonese large language model (LLM) as akin to training a model from scratch. This involves taking an existing open-source model, such as Meta’s Llama or Alibaba’s Qwen, and augmenting it with Cantonese-specific data. Votee sources this data through various means, including online scraping from platforms like Radio Television Hong Kong (RTHK), contributions from the community and universities, and synthetic data generation. These efforts have expanded the corpus of Cantonese data from 100 million tokens to over 500 million.

Votee AI’s models are designed with approximately 70 billion parameters, which is considerably smaller than leading models currently available. Despite this, Ting asserts that their models are sufficiently capable of understanding and reasoning in Cantonese. The training costs, while not insignificant, are also more manageable compared to those incurred by larger AI laboratories. Ting estimates that Votee utilizes between 500 million and 1 billion tokens for training, as opposed to the trillions needed for English-language models, with an estimated cost of around $250,000.

A Broader Movement Towards Language Inclusion

The endeavor to develop models for low-resource languages is gaining traction globally. For instance, Indosat, Indonesia’s second-largest telecommunications company, is working on Sahabat AI, an open-source model focused on Indonesian languages. Similarly, AI Singapore is developing SEA-LION, which encompasses 11 under-resourced Southeast Asian languages. South Korea has taken a more aggressive approach by organizing a state-sponsored competition, dubbed the “AI Squid Game,” to identify national champions for homegrown foundation models, supported by a substantial AI budget of approximately $6.8 billion for 2026.

This movement aligns with the concept of “sovereign AI,” which advocates for countries and companies to maintain ownership of their data, models, and infrastructure instead of relying on foreign entities. Ting notes the increasing necessity for AI, emphasizing that countries do not want to be dependent on external sources that could potentially restrict access. He acknowledges the complexity of fully realizing the sovereign AI vision, suggesting that nations should focus on owning foundational models and the applications built upon them.

Governments may not require models as advanced as those at the forefront to automate basic tasks. Ting explains that even smaller models, with as few as 1 billion parameters, can fulfill such needs. For more sophisticated capabilities, governments could leverage the outputs of powerful English- or Mandarin-language models, adapting them for local languages.

Votee collaborates with various models, including those from MiniMax and SenseTime, and utilizes Nvidia chips in its operations. Ting likens the company’s position to that of a high school student who maintains friendships across different social groups, allowing for a diverse range of technological partnerships.

While Votee is committed to preserving linguistic diversity, it operates as a for-profit entity. Its primary clientele consists of governments and corporations, and Ting claims that the company is profitable, with revenues exceeding operational costs, largely supported by client contracts. Notably, Allan Zeman, a prominent figure in Hong Kong’s business landscape, has joined Votee as an advisor.

Votee’s aspirations extend beyond Hong Kong. Ting reveals that the startup is engaged in “active discussions” with AI Singapore and plans to expand its reach throughout Southeast Asia. Additionally, there are ambitions to leverage AI for the preservation of endangered languages across various regions, including East Asia, North America, and Africa.

Ting likens the impact of English-language AI on other languages to a “typewriter moment,” where the productivity gains associated with English lead to a decline in the use of native languages. He expresses a desire to provide other languages with a fighting chance in a landscape dominated by English and Mandarin AI. "Every language that dies, you lose another way of seeing the world," he states, emphasizing the importance of linguistic diversity in understanding different perspectives.

Pak-Sun Ting is scheduled to speak at the Fortune Leaders Forum on September 8 in Macau, where he will discuss these pressing issues further. The forum aims to bring together Fortune 500 executives and founders of leading Asian companies to shape the future of leadership amid evolving technological landscapes.