Hong Kong's HKGAI launches DeepSeek V4-based model for domestic chips
Synopsis
Hong Kong's government-backed HKGAI has launched HKGAI-V3, a DeepSeek V4-based model that runs on domestic chips including Huawei's Ascend C, achieving a near-hundredfold boost in agent runtime and a 28-hour uninterrupted autonomous session — a direct counter to US semiconductor export controls.
Key Takeaways
HKGAI launched HKGAI-V3 on 3 June 2026 , a large language model built on DeepSeek V4 and optimised for domestic chips.
The model delivers a tenfold improvement in token compression efficiency and an almost hundredfold increase in uninterrupted agent runtime.
Agent Workshop , built on HKGAI-V3 , ran stably for up to 28 hours without interruption in a single session, according to director Guo Yike .
Partners include Huawei ( Ascend C chips), Inspur Group , China Unicom Global , China Telecom Global , and China Mobile International .
HKGAI was established in October 2023 under the InnoHK programme and also operates HKChat , Hong Kong 's first Cantonese-enabled chatbot.
The centre's stated goals include commercialising its products and exporting Chinese AI technology internationally.
Hong Kong's Generative AI Research and Development Centre (HKGAI) has unveiled HKGAI-V3, a large language model built on DeepSeek V4 engineered to run on domestic chips, marking a significant step in the government-backed lab's push to commercialise homegrown AI and expand Chinese AI's global footprint. The launch was announced on Wednesday, 3 June 2026, at an official launch event in Hong Kong.
What HKGAI-V3 delivers
The new model delivers a tenfold improvement in token compression efficiency, reducing the number of tokens a model must process and thereby cutting response times, according to the centre. It also achieves an almost hundredfold increase in uninterrupted agent runtime compared to its predecessor, enabling far more complex autonomous task execution.
The centre said these gains represent 'significant improvements' in both efficiency and agentic capabilities, positioning HKGAI-V3 as a competitive foundation for enterprise and government AI workloads.
Agent Workshop: 28-hour autonomous sessions
HKGAI director Guo Yike demonstrated Agent Workshop, an AI agent platform built on top of HKGAI-V3, at the launch event. The platform operated stably for up to 28 hours without interruption in a single session to produce a research report, according to Guo Yike.
Agent Workshop represents the centre's first commercially oriented product built on its own model stack, signalling a shift from pure research toward deployable AI services.
Domestic chip compatibility and strategic context
A defining feature of HKGAI-V3 is its ability to run on domestic chips — a capability with clear strategic significance as China navigates tightening US export controls on advanced semiconductors. The centre's partners include Huawei, whose Ascend C chip architecture is among the primary domestic alternatives to restricted Nvidia hardware, alongside Inspur Group, China Unicom Global, China Telecom Global, and China Mobile International.
HKGAI was established in October 2023 under the Hong Kong government's InnoHK funding programme, with a mandate to develop AI tailored to local needs. Its existing products include HKChat, the city's first Cantonese-enabled chatbot trained on local services, laws, and regulations, as well as productivity tools deployed across government departments. The centre is affiliated with Hong Kong University of Science and Technology (HKUST).
Why it matters
The launch underscores Hong Kong's ambition to serve as a conduit for exporting Chinese AI technology internationally, leveraging its unique regulatory and financial position. By grounding HKGAI-V3 in DeepSeek V4 — one of China's most capable open-weight model lineages — and optimising it for domestic silicon, the centre is building a stack that is insulated from Western chip restrictions.
The involvement of the three major Chinese state telecoms operators as partners also points to a commercialisation pathway through cloud and enterprise channels beyond Hong Kong's borders.
What's next
With Agent Workshop now live and the model validated on domestic hardware, HKGAI is expected to accelerate enterprise deployments and pursue international licensing or partnership arrangements. The centre's next test will be whether HKGAI-V3's efficiency gains translate into cost-competitive services against globally available models — a benchmark that will determine how far its export ambitions can reach.
Point of View
Hong Kong is demonstrating that the Chinese AI ecosystem can function independently of Nvidia hardware — exactly the proof-of-concept Beijing needs as export controls tighten. What mainstream coverage tends to underplay is Hong Kong's structural role as a soft-power relay: its common-law framework and international financial ties give HKGAI-built products a credibility and market-access advantage that purely mainland Chinese labs lack when approaching overseas enterprise clients. The near-hundredfold agent runtime improvement also signals that the real battleground in 2026 is not raw benchmark scores but long-horizon autonomous task execution — a dimension where Western frontier labs have not yet set a definitive standard. Investors and policymakers watching the chip-war playbook should note that the state telecoms trio's involvement as distribution partners gives HKGAI a ready-made global sales channel through existing enterprise and cloud relationships.
NationPress
21 Jul 2026
Frequently Asked Questions
What is HKGAI-V3 and when was it launched?
HKGAI-V3 is a large language model developed by Hong Kong's Generative AI Research and Development Centre, built on DeepSeek V4 and designed to run on domestic chips. It was officially launched on 3 June 2026 at an event in Hong Kong, where the centre highlighted its improved token compression efficiency and extended autonomous agent runtime.
What chips does HKGAI-V3 run on?
HKGAI-V3 is engineered to run on domestic chips, with Huawei's Ascend C architecture among the primary hardware platforms. This design choice is strategically significant given ongoing US export restrictions on advanced semiconductors such as Nvidia GPUs that limit access for Chinese AI developers.
What is Agent Workshop and what can it do?
Agent Workshop is an AI agent platform built by HKGAI on top of the HKGAI-V3 model. According to HKGAI director Guo Yike, it operated stably for up to 28 hours without interruption in a single session to produce a research report, demonstrating the model's near-hundredfold improvement in uninterrupted agent runtime.
Who are HKGAI's partners for this model?
HKGAI's partners include Huawei (Ascend C chips), Inspur Group, China Unicom Global, China Telecom Global, and China Mobile International. The involvement of the three major Chinese state telecoms operators suggests a commercialisation strategy that extends well beyond Hong Kong's domestic market.
How does HKGAI-V3 fit into the broader China AI and chip landscape?
HKGAI-V3 represents a concrete example of the Chinese AI ecosystem adapting to Western chip restrictions by validating a competitive model on domestic silicon. Built on DeepSeek V4 — one of China's leading open-weight model lineages — and backed by state telecoms distribution partners, it positions Hong Kong as a potential international gateway for Chinese AI technology.
