Hong Kong's HKGAI-V3 model runs on domestic chips with 10x efficiency

The Hong Kong Generative AI Research and Development Centre (HKGAI) officially released HKGAI-V3, a large language model built on DeepSeek V4 and optimized to run on domestic semiconductor chips. The government-backed lab announced the model at a launch event on Wednesday, highlighting significant gains in efficiency and agentic capabilities. HKGAI-V3 delivers over tenfold improvement in token compression, reducing the number of tokens a model needs to process and thereby cutting response times. It also achieves an almost hundredfold increase in uninterrupted agent runtime, enabling complex multi-step tasks without manual intervention.

A key showcase is Agent Workshop, an AI agent platform built on HKGAI-V3. In a single session, it operated stably for up to 28 hours to produce a comprehensive research report. According to HKGAI director Guo Yike, this demonstrates the model's ability to handle long-horizon tasks autonomously. The centre also operates HKChat, Hong Kong's first chatbot trained on local services, laws, and regulations with Cantonese language support. HKGAI was established in October 2023 under the government's InnoHK funding programme, focusing on AI models tailored to Hong Kong's needs.

The launch underscores Hong Kong's push to commercialize homegrown AI and export Chinese technology overseas. By ensuring compatibility with domestic chips, HKGAI-V3 reduces dependence on foreign hardware, addressing geopolitical concerns. The model's enhanced agent capabilities position it as a practical tool for government and enterprise use, from research automation to public service chatbots.