Hong Kong Is Developing a DeepSeek‑Based AI Model That Runs on Domestic Chips

Hong Kong is developing a new large language model based on DeepSeek technology that can run entirely on China‑made AI chips, marking a significant step in the city’s effort to build a locally controlled AI ecosystem.

The initiative is led by the Hong Kong Generative AI Research and Development Centre (HKGAI) and reflects a broader strategy to strengthen regional AI capabilities while reducing dependence on foreign hardware and cloud providers.

A Local AI Model Built on DeepSeek

Researchers at HKGAI are building the new system on top of DeepSeek’s architecture, adapting it so the model can operate fully on domestically produced chips rather than relying on U.S.-designed GPU platforms.

The upcoming model, reportedly called HKGAI‑V3, is expected to be introduced as part of Hong Kong’s effort to create deployable AI infrastructure for government, business, and potential overseas users.

This approach allows developers to:

• Run AI workloads locally without depending on foreign cloud services
• Maintain tighter control over data security and governance
• Support the growth of regional chip and software ecosystems

The compatibility with China‑made hardware is central to the strategy. Domestic chip technologies—particularly those developed by companies such as Huawei—are increasingly being adapted for advanced AI workloads.

Part of Hong Kong’s “Sovereign AI” Strategy

The project sits within a wider push toward what policymakers and researchers describe as sovereign AI—AI systems developed, trained, and deployed locally.

This model prioritizes:

• Control over training data and infrastructure
• Reduced dependence on foreign technologies
• Support for local language and cultural contexts

Hong Kong’s AI ecosystem is also designed to handle trilingual usage—Cantonese, Mandarin, and English—reflecting the city’s linguistic environment and regional business role.

Earlier Milestone: HKGAI V1

The new chip‑compatible model builds on earlier work from the same research center.

In 2025, the Hong Kong government confirmed that HKGAI V1, the city’s first homegrown large language model, was developed through full‑parameter fine‑tuning of DeepSeek models combined with continued training on local datasets.

That system represented a major step toward building a domestic AI platform capable of serving public agencies and businesses within the region.

Why Chip Compatibility Matters

Adapting AI models to run on domestic chips has become a strategic priority across China’s technology sector.

Recent DeepSeek releases have already been optimized for a mix of hardware platforms, including domestic processors, signaling a broader industry move away from exclusive reliance on Nvidia GPUs.

This shift has ripple effects across the semiconductor market. When DeepSeek announced newer models designed with greater compatibility for Chinese hardware, semiconductor stocks in China and Hong Kong rose on expectations of increased demand for locally produced chips.

What Comes Next

If the HKGAI‑V3 system launches as planned, Hong Kong could position itself as a regional hub for deployable AI models built on domestic infrastructure.

Such systems could be used across government services, local enterprises, and international markets looking for alternatives to Western AI stacks. The strategy also aligns with broader regional trends toward building AI systems that can operate independently of foreign computing platforms.

While many technical details of the upcoming model remain limited, the direction is clear: Hong Kong is moving toward an AI stack that combines DeepSeek software architecture with China‑developed hardware, aiming to strengthen technological autonomy while expanding its role in the global AI ecosystem.