Hong Kong on Wednesday unveiled the latest version of its homegrown large language model, HKGAI V3, marking a major step in efforts to build artificial intelligence tools tailored to the city's industry scenarios and linguistic context.

Developed by the Hong Kong Generative AI Research and Development Center (HKGAI), the new model delivers improvements in operating efficiency and agent performance. Token compression efficiency has improved more than tenfold, while uninterrupted agent runtime has increased nearly a hundredfold compared with the previous version.

HKGAI V3's Agent Workshop, the core platform for Hong Kong's first productivity-grade "super agent", can run continuously for up to 28 hours in a single session.

Developers said that infused with Hong Kong's "cultural DNA", the model can interpret inputs and express ideas in ways that better reflect local users' habits.

Secretary for Innovation, Technology and Industry Sun Dong said, "With the release of the V3 model, Hong Kong is not sitting on the sidelines of the global wave of technological innovation driven by AI, but is gradually moving toward the forefront."

Sun said he expects the model to enable more practical, locally relevant vertical applications, including enhancements to HKChat and HKPilot — the city's AI chatbot and AI-powered document processing tool — allowing users in Hong Kong to enjoy higher-quality, more culturally attuned AI agent services.

Sun added that the Hong Kong Special Administrative Region government is taking a multipronged approach to advance both "the industrialization of AI" and "the integration of AI into traditional industries", as part of efforts to strengthen the city's AI ecosystem.

"Pretraining a large language model is like raising a child," said Guo Yike, director of HKGAI and provost of the Hong Kong University of Science and Technology. "No matter how capable it is in general knowledge or how vast its parameters are, it cannot truly understand boundaries and context until it is aligned with local regulations and societal values."

Guo added that technical benchmarks alone are no longer sufficient to evaluate AI models. "What truly makes a difference is localization and security compliance," he said.

Developers said the model can be used by government bureaus, businesses and the public. Within its Agent Workshop, for instance, AI agents can autonomously break down user instructions into tasks, call on external tools and collaborate across platforms to complete workflows.

In addition, the model will be open-sourced under the name ClawNet, providing easy access for enterprises and research institutions to develop customized AI agents.

HKGAI V1, released in February 2025, was Hong Kong's first homegrown large language model.

irisli@chinadailyhk.com