Hong Kong unveils HKGAI-V3 model built on domestic tech and DeepSeek V4

The Hong Kong Generative AI Research and Development Centre (HKGAI) has officially launched a new DeepSeek-based large language model that can run on domestic chips, as the government-backed lab seeks to commercialise its products and export Chinese AI overseas.

The HKGAI-V3 model, built on DeepSeek V4, has achieved “significant improvements” in efficiency and agentic capabilities, the centre said on Wednesday.

The home-grown model delivered over tenfold improvement in the efficiency of token compression, which reduced the number of tokens a model needs to process and therefore its response time, and an almost hundredfold increase in uninterrupted agent runtime, according to the lab.

Agent Workshop, an AI agent platform the HKGAI developed based on HKGAI V3, for instance, operated stably for up to 28 hours without interruption in a single session to produce a research report, HKGAI director Guo Yike said on Wednesday at the launch event.

HKGAI, established in October 2023 under the government’s InnoHK funding programme, focuses on developing AI models and products tailored to Hong Kong’s needs. It operates HKChat, the city’s first Cantonese-enabled chatbot trained to offer information on local services, laws and regulations, and several AI productivity tools for government departments.