๐ญ๐ฐSCMP TechnologyโขStalecollected in 2m
Hong Kong launches DeepSeek-based model for domestic chips
๐กFirst DeepSeek-based model optimized for domestic chips, offering a path to bypass high-end GPU supply constraints.
โก 30-Second TL;DR
What Changed
HKGAI-V3 is built upon the DeepSeek V4 architecture.
Why It Matters
This development signals a strategic shift toward hardware-software co-optimization in the Chinese AI ecosystem, reducing reliance on foreign GPU architectures.
What To Do Next
Evaluate the HKGAI-V3 model if you are developing AI applications for deployment on domestic Chinese hardware infrastructure.
Who should care:Developers & AI Engineers
Key Points
- โขHKGAI-V3 is built upon the DeepSeek V4 architecture.
- โขThe model is specifically optimized to run on domestic Chinese AI chips.
- โขAchieved a tenfold improvement in token efficiency and enhanced agentic capabilities.
๐ง Deep Insight
Web-grounded analysis with 15 cited sources.
๐ Enhanced Key Takeaways
- โขHKGAI-V3 is designed with embedded local cultural context and supports Cantonese, alongside Chinese and English, to improve comprehension and outputs for city-specific use cases in Hong Kong.
- โขThe model's Agent Workshop platform demonstrated a near hundredfold increase in uninterrupted agent runtime, operating stably for up to 28 hours in a single session for tasks such as generating research reports.
- โขHKGAI-V3 is optimized for both mainstream and domestic Chinese hardware, specifically mentioning compatibility with Huawei Technologies' Ascend 910C chips.
- โขThis initiative is a key component of Hong Kong's broader 'sovereign AI' strategy, which aims for the jurisdiction to develop and operate AI systems on its own infrastructure to ensure security and data sovereignty.
- โขThe Hong Kong Generative AI Research and Development Centre (HKGAI Centre), established in October 2023 under the InnoHK programme, previously developed HKChat, the city's first Cantonese-enabled chatbot for local services and regulations.
๐ ๏ธ Technical Deep Dive
- DeepSeek V4 is a Mixture-of-Experts (MoE) model, with the Pro version featuring 1.6 trillion total parameters (49 billion active) and the Flash version having 284 billion total parameters (13 billion active).
- It supports an ultra-long context window of up to 1 million tokens, equivalent to 15-20 full novels.
- Key architectural innovations include a Hybrid Attention Architecture (combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA)) for enhanced long-context efficiency.
- The model incorporates Manifold-Constrained Hyper-Connections (mHC) to stabilize signal propagation across deep layers and utilizes the Muon Optimizer for faster convergence and improved training stability.
- DeepSeek V4 employs mixed precision training, with MoE expert parameters using FP4 precision and most other parameters using FP8 to maximize memory efficiency.
- The V4 model has been validated across both Nvidia Graphics Processing Units (GPUs) and Huawei Ascend Neural Processing Units (NPUs), indicating a fine-grained expert parallelism scheme that supports both platforms.
- HKGAI-V3 involves "full-parameter fine tuning for localisation" of the DeepSeek V4 architecture.
- DeepSeek V4 also features native multimodal support, processing text, images, video, and audio from scratch.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
Hong Kong will strengthen its position as a regional AI innovation hub, particularly in 'sovereign AI'.
The HKGAI-V3 initiative, with its focus on domestic hardware optimization and local cultural context, aligns with Hong Kong's strategy to build independent AI capabilities and foster a tailored AI ecosystem.
The adoption of HKGAI-V3 will accelerate the commercialization and deployment of AI in Hong Kong's public services and enterprises.
The model's enhanced efficiency, agentic capabilities, and localization (Cantonese support, local cultural context) are designed to improve comprehension and outputs for city-specific use cases, including government departments and local services.
China's domestic AI chip industry will experience increased demand and development due to initiatives like HKGAI-V3.
The explicit optimization of HKGAI-V3 for domestic Chinese AI chips, such as Huawei's Ascend 910C, directly supports the broader national goal of reducing reliance on foreign semiconductor technologies.
โณ Timeline
2023-07
DeepSeek was founded by Liang Wenfeng, funded by High-Flyer.
2023-10
Hong Kong Generative AI Research and Development Centre (HKGAI Centre) was established under the InnoHK programme.
2024-12
DeepSeek-V3-Base and DeepSeek-V3 (chat) models were released.
2025-01
DeepSeek-R1 model and its eponymous chatbot launched, gaining international prominence.
2025-02
HKGAI-V1, Hong Kong's first homegrown large language model, was released.
2026-04-24
DeepSeek V4 (Pro and Flash versions) was officially launched as a preview and open-sourced.
2026-06-03
Hong Kong Generative AI Research and Development Centre officially unveiled HKGAI-V3.
๐ Sources (15)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: SCMP Technology โ