HKGAI V1, a DeepSeek 671B fine-tune, now carries HKChat’s 700,000 users and 100-plus government pilots; the 2026 test is workflow, not weights
By: Robert Sterling – SeaPRwire – In November 1998 the Hong Kong government published a document called Digital 21, a strategy for making the territory a leading digital city. Four months later the Financial Secretary stood up and announced Cyberport, a campus on the southern shore of Hong Kong Island that would host a cluster of technology firms and, in the phrasing of the time, put the city on the map for the coming century. The first tenants moved in during 2002. By the time the second phase opened, the dot-com crash had already taken most of the companies that were supposed to fill it.
Version two of that ambition sits in the same compound. In December 2024, Cyberport’s Artificial Intelligence Supercomputing Centre began serving customers with roughly 1,300 petaflops of computing power. Sixteen months later, in February 2025, a research centre led by the Hong Kong University of Science and Technology released a large language model that handles Cantonese, Mandarin and English in the same conversation. The centre is called HKGAI. The model is called HKGAI V1. It was built, its team says, in about a month.
Those two facts belong in the same paragraph because they explain each other. The twenty-six years between them were not wasted, but they were not spent on the thing that finally mattered. What changed in 2025 was not local talent, local capital or local ambition. What changed was that someone else had already paid for the expensive part.
A campus built for something else
Cyberport was sold to the public as a property project with a technology policy attached, which is roughly what it became. The criticisms of 1999 have not aged badly: the land was rezoned from residential, the developer was chosen without an open tender, and the residential portion turned out to be the part with reliable cash flow. For a decade the campus hosted startups, a digital entertainment cluster and a handful of multinational back offices, none of which required a supercomputer.
The 2017 Smart City Blueprint added sensors and open data. The 2023 policy address added a push on artificial intelligence. None of it changed the underlying arithmetic, which was that training a frontier model cost more than any Hong Kong institution could justify spending on a bet with no obvious customer.
That arithmetic collapsed in 2024 and 2025 in a way that is easy to miss if you read technology news only in English. Chinese labs began releasing model weights that anyone could download and fine-tune. DeepSeek published R1 early in 2025. Within weeks, a Hong Kong team that had been assembling data, evaluation sets and legal corpora for eighteen months had something to attach them to.
One month, one base model
The published description of HKGAI V1 is unusually specific for a launch announcement: a full-parameter fine-tune of the DeepSeek 671B model, adapted for Cantonese, Mandarin and English, with local knowledge bases and government open data folded in. The full-parameter detail matters. LoRA adapters are cheap and reversible; retraining every parameter of a 671-billion-parameter model is not. It signals a decision to own the resulting artefact rather than rent it.
The time-to-release number is the interesting one. Eighteen months of corpus work, evaluation design and legal review, followed by roughly a month of training once the base model existed. That ratio, eighteen to one, is the actual news of 2025 for every mid-sized jurisdiction that has been told it cannot afford artificial intelligence. The expensive half of the work has been socialised by someone else. The remaining half, which is data, evaluation and domain adaptation, is the half a city like Hong Kong can realistically do.
Whether that division of labour is comfortable is a separate question. A model that is a fine-tune inherits the priors, the refusals and the failure modes of its base. Hong Kong’s version can speak Cantonese; it cannot decide what Cantonese public life should look like in ten years, because that decision was made upstream, in a lab in Hangzhou, in a set of training choices nobody in Cyberport saw.
Cantonese is not a small language, it is a poorly written-down one
The reason generic models handle Cantonese badly is not that the language lacks speakers. Estimates put the number somewhere above eighty million, concentrated in Guangdong, Hong Kong, Macau and overseas communities that have been exporting films, songs and restaurant menus for a century. By speaker count, Cantonese is larger than Italian and roughly the size of Korean.
Digital resources do not follow speaker counts. They follow institutional habits: whether a language is used in government correspondence, whether it has a standard written form, whether decades of transcribed speech exist in a machine-readable corpus. Cantonese has an unusual problem on all three counts. Official Hong Kong Chinese is standard written Chinese, which is grammatically Mandarin. Spoken Cantonese has its own vocabulary, its own particles and a writing convention that appears mostly in chat messages, comics and advertising.
The practical consequence for a model is code-switching. A real Hong Kong sentence about a lease renewal will start in Cantonese, drop into an English term for the legal instrument, and finish with a Mandarin-flavoured written phrase because that is how the form was drafted. Getting that mix right is not a translation task. It is a register task, and registers are learned from transcribed speech, not scraped from news sites.
This is where eighteen months of corpus work went. The team’s public statements describe news interviews and everyday dialogue as the base material, with rolling correction of errors. Anyone who has built a speech corpus knows what that actually costs: consent forms, transcription guidelines, a second pass by a different transcriber, and a policy for the inevitable recording where two people talk over each other for forty seconds.
Compute, counted
Hong Kong’s answer to the compute question has been to build, subsidise and then build again. The numbers below are the ones a reader can verify from government papers and Legislative Council briefings, and they tell a story about sequencing rather than scale.
| Layer | What exists | When | Note |
| Supercomputing | 1,300 PFLOPS at Cyberport | December 2024 | Ramped to 3,000 PFLOPS during 2025 |
| Subsidy | HK$3 billion, three-year AI Subsidy Scheme | October 2024 | Ten projects approved by end-April 2025, more than 30 later |
| Institution | HK$1 billion AI Research Institute | Planned for 2026-27 | Still being designed by the Digital Policy Office |
| Training | HK$50 million universal AI training | 2026-27 budget | 200-plus events, 50,000 participants targeted |
| Pipeline | Sandu data park, 180,000 PFLOPS | Target 2032 | 11 hectares, awarded on a double-envelope basis |
| Private | HKSTP and SenseTime, 40,000 PFLOPS | Three phases to 2030 | Described as the largest domestic-chip compute centre locally |
Read the column on the right and a pattern appears. Each announcement is larger than the last and further out in time. The 3,000 petaflops available now is real and about 80% utilised, according to government statements, which is a healthy number for a subsidised facility. The 180,000 petaflops in 2032 is a planning figure, and planning figures in Hong Kong have a way of arriving late and smaller.
The subsidy scheme is the more revealing instrument. Ten approved projects by the end of April 2025, against a HK$3 billion envelope, works out to an average commitment far below what the headline number implies. A scheme that approves ten projects in six months is not short of money. It is short of applicants who can use a thousand petaflops without wasting them.
Five apps, one base model
HKGAI’s public product list is short and deliberately unglamorous. HKChat answers citizen questions in Cantonese about transport, weather, benefits and legal aid. HKPilot drafts and edits official documents. HKMeeting turns a mixed-language meeting into minutes. Lexi HK searches Hong Kong judgments and legislation. HKEnviro assembles environmental assessment material.
The distribution numbers are the ones to hold on to. HKChat passed 700,000 registered users by early 2026, up from roughly 160,000 at the end of 2025. The suite has been piloted across more than a hundred government departments and offices. Lexi HK’s retrieval layer is built on a set of 58,130 court judgments, a corpus that no commercial vendor would bother assembling for a market of this size.
That last point is the strongest argument for a publicly funded local model, and it has nothing to do with national pride. The commercial incentive to build a Cantonese legal retrieval system for seven million people is close to zero. The social value of one is high, because the alternative is that residents and small firms continue to navigate a bilingual legal system with a search engine and a phone call.
The counter-argument is equally concrete. A government-subsidised assistant that 700,000 people have registered for and an unknown number use weekly is a service, not a transformation. Registration counts are the easiest metric in the world to inflate and the hardest to interpret.
Two small jurisdictions, two bets
Singapore is the obvious comparison, and it took a different route with the same raw material. AI Singapore’s SEA-LION family, first released in 2023, was built as a set of smaller open models tuned for the languages of Southeast Asia, with an explicit emphasis on Bahasa Indonesia, Thai, Vietnamese and Tamil alongside English. The strategy was to be useful at eight billion parameters rather than impressive at six hundred.
Hong Kong went the other way: one very large base, one deep adaptation, five applications aimed at its own residents. The trade-off is legible. A 671-billion-parameter fine-tune can hold legal nuance and code-switching in a way a small model cannot, but it costs real money to serve and it inherits more of its base model’s personality. A family of small models is cheap to run and easier to audit, and it will lose an argument about a 1970s tenancy clause.
Which is right depends on what the model is for. If the purpose is to answer a resident’s question about a benefit application, the smaller model is adequate and the cost per query matters. If the purpose is to draft a bilingual submission that a court might later read, the larger one earns its electricity. Hong Kong’s choice reveals which use case it weighted first: the legal corpus, the judgment retrieval and the official document drafting, not the consumer chat.
The 2026 direction suggests a partial correction. HKGAI’s public roadmap now talks about agents and workflow tools rather than raw model size, and its partner list has grown to include Ant Group, Lenovo and a local cloud operator, with roughly thirty firms signed up for joint deployment. Partners are a leading indicator of a different kind: they appear when someone believes there is a paying customer at the end.
Eighty-eight per cent, and fifty-four
Two surveys published within four months of each other describe the same market from different heights, and the gap between them is where the next two years will be decided.
The Hong Kong Productivity Council interviewed about 800 companies in September 2025 and released the results in October. Eighty-eight per cent of employees in the surveyed firms had used AI tools at work, mostly for customer service, data analysis and marketing. Ninety-two per cent of respondents planned to fold AI into formal workflows, and 24% intended to do it within a year. So far, so enthusiastic.
Then the governance numbers. Forty-five per cent had an officially recognised AI platform for staff to use. Fifty-four per cent said they had no complete or ongoing AI governance framework, and no policy to speak of. Five challenges were named in order: lack of expertise, data privacy, integration with existing systems, employee resistance, and cost.
The government’s own answer to that gap arrived in April 2025 as the Hong Kong Generative AI Technical and Application Guideline, drafted by HKGAI on commission from the Digital Policy Office. Read as a document rather than an announcement, it is the first attempt to put the territory’s position on model risk into writing: data handling, human review of outputs, disclosure, and a preference for locally hosted models in official work. Its scope is government departments. The 54% above describes everybody else.
Now drop a level. The Standard Chartered Hong Kong SME Leading Business Index for the first quarter of 2026 surveyed 819 small and medium enterprises and asked a narrower question. Fifty-five per cent were using AI or planned to start within a year. That figure includes both, which is the sort of rounding that makes a market look further along than it is.
Inside the group that already uses AI, the shape of adoption is thin. Chatbots 69%, text and document generators or OCR 55%, voice and speech-to-text 35%, image and video 34%, data analysis and prediction 18%. Fifty-nine per cent use more than one category, which sounds like depth until you notice that the categories are all consumer tools with a login page.
The number that should worry a policy office is smaller: 27% of AI-using SMEs planned to increase their AI spending in the coming year. Seventy-five per cent said they had expanded usage in 2025. Usage up, budget flat. That is what a market looks like when the tool is free or nearly free at the margin, and when nobody has yet connected it to a line item that a manager is measured on.
What small firms bought
Walk through the arithmetic of a thirty-person trading company in Kwun Tong, the kind of firm that makes up the bulk of the 1.7 million small and medium enterprises in the territory. It has a bookkeeper, a sales manager who handles overseas buyers, a shipping clerk and a handful of warehouse staff. The owner has an accountant, a bank relationship and a compliance obligation that grows every year.
What has that firm bought in the past two years? Almost certainly a subscription to a chatbot, used by two or three people, billed monthly in US dollars. Possibly an OCR tool for invoices, which works well until a supplier sends a photographed delivery note at an angle. Probably nothing that touches inventory, customs declarations or payroll.
Why not? Because the tools that touch those functions are the ones that require an integration project, and an integration project requires someone who can specify it. The Productivity Council’s survey names this as the top obstacle, ahead of money. A firm with no IT manager cannot buy its way out of not knowing what to ask for.
This is also where the subsidy schemes run into their own design. Most Hong Kong technology vouchers reimburse expenditure after approval, which is fine for a company with a finance function and a cash buffer, and a barrier for a firm that would have to fund the whole project first. The programmes that work for small firms tend to be the ones where a vendor is pre-approved and the paperwork is done by the vendor. Nothing about that is specific to Hong Kong.
The part of 1999 that repeats
The temptation, reading the 2026 numbers, is to declare that the second attempt has learned from the first. There is evidence for it. The compute is real, the model exists, the applications are in production, and the subsidy is being disbursed rather than merely announced.
There is also a pattern worth naming. In 1999 the assumption was that if you built the campus, the companies would come. In 2026 the assumption is that if you build the compute and the model, the workflows will follow. Both assumptions put infrastructure ahead of the unglamorous middle layer, which is the part where an organisation decides to change how it does something and accepts the temporary loss of productivity that comes with it.
Hong Kong’s advantage in this cycle is not the model. It is the density of the problem. A city with one legal system, one tax regime, one language mix and a few hundred thousand firms is a small enough test bed to actually finish an integration, which is more than can be said for jurisdictions trying to do the same thing across twenty-seven member states.
Its disadvantage is the same density read the other way. Every integration in Hong Kong is a bespoke integration, because the firms are small and their processes are idiosyncratic. There is no equivalent of a national health service or a nationwide retailer whose adoption of a system drags ten thousand suppliers with it.
Three numbers worth tracking into 2027
Ignore parameter counts and ribbon-cuttings. Three indicators will tell you whether the 2025 model release turned into something durable.
- The share of HKPC-surveyed firms with a written AI governance policy. It was 46% implied in late 2025, which is the ceiling on any serious deployment.
- The number of AI-using SMEs that raise their AI budget two years in a row. Twenty-seven per cent intended to raise it once; the second year is the one that separates habit from experiment.
- Approved projects under the AI Subsidy Scheme, measured against compute actually consumed rather than compute available. Ten approvals against 1,300 petaflops is the ratio to watch.
Weights are not the constraint any more
The most useful thing about February 2025 is how quickly it made an old argument obsolete. For a decade, the reason given for Hong Kong’s thin record in deep technology was that the city could not build the base layer: no chip industry, no frontier lab, no appetite for a decade of losses. That argument is now only partly true, because the base layer became available as a download.
What is left is the work that cannot be downloaded. Transcribing Cantonese speech until the model stops guessing. Writing the governance policy that 54% of firms do not have. Sitting with a thirty-person trading company for three weeks to work out which of its twelve weekly processes should be handed to a machine, and which should be abolished instead. None of that is a research problem, and none of it will appear in a petaflop announcement.
The 1999 version of this story ended with a property development and a lesson about sequencing. The 2026 version has better raw material and the same middle layer to cross. If the next five years produce another 3,000 petaflops and no change in the number of firms that can describe their own AI workflow in a sentence, the lesson will simply have moved up a level.