Hong Kong launches DeepSeek-based AI model designed to run on domestic chips
HKGAI launched HKGAI-V3, a DeepSeek-based model that runs on domestic chips. The lab says it improves token compression efficiency and agent runtime significantly.
Intelligence analysis by GPT-5.4 Mini
Hong Kong’s government-backed AI lab has unveiled HKGAI-V3, a large language model built on DeepSeek V4 and designed to run on domestic chips. The launch is framed as a step toward commercialising local AI products and exporting Chinese AI overseas.
Hong Kong made a new AI model like a smart robot brain that can work on local computer chips. The lab says it answers faster and can keep doing tasks for a very long time, like a worker who stays on the job all night.
Analysis
What launched
The Hong Kong Generative AI Research and Development Centre, or HKGAI, has released HKGAI-V3, a large language model based on DeepSeek V4. The centre says the model is designed to run on domestic chips, which matters because hardware compatibility can affect where and how AI systems are deployed.
What the lab claims
HKGAI said the new model delivered more than a tenfold gain in token-compression efficiency, which means it can reduce the amount of text a model has to process and therefore answer faster. The centre also said it saw an almost hundredfold increase in uninterrupted agent runtime.
At the launch event, HKGAI director Guo Yike said the lab’s Agent Workshop platform, built on HKGAI-V3, was able to run stably for up to 28 hours in one session while producing a research report.
Why Hong Kong is doing this
HKGAI was established in October 2023 under the government’s InnoHK funding programme. Its mission is to develop AI models and products suited to Hong Kong’s needs. The centre already operates HKChat, described as the city’s first Cantonese-enabled chatbot trained to provide information on local services, laws and regulations, and it also makes AI productivity tools for government departments.
The launch suggests HKGAI wants to move beyond research and turn its models into products that can be used locally and sold more broadly, including overseas markets. The article presents that as part of a broader push to commercialise Chinese AI.
Key points
- HKGAI launched HKGAI-V3, a DeepSeek-based large language model designed to run on domestic chips.
- The centre says the model improved token compression efficiency by more than ten times.
- HKGAI said uninterrupted agent runtime increased by almost one hundred times.
- Its Agent Workshop platform reportedly ran for up to 28 hours in one session to produce a research report.
- HKGAI says its goal is to commercialise products and export Chinese AI overseas.
If the performance claims hold up, HKGAI-V3 could make AI tools faster and easier to deploy on domestic hardware. That would support Hong Kong’s push to build local AI products and sell them beyond the city.
The article only reports the lab’s own claims, so real-world performance may differ outside launch demos. If the model depends on narrow conditions, the promise of running well on domestic chips and sustaining long agent tasks may be harder to reproduce at scale.
