At the 2026 World Artificial Intelligence Conference (WAIC) Meteorological Session, the China Meteorological Administration (CMA) unveiled “Fenghe,” an AI-powered large language model designed specifically for meteorological services, and launched its global open-source initiative.
As China’s first meteorological service domain model with hundreds of billions of parameters, Fenghe is expected to accelerate the integration of artificial intelligence into the full chain of meteorological services and help transform traditional weather services into more intelligent, efficient and personalized systems.
Developed by the CMA Public Meteorological Service Centre in collaboration with the Xiong’an Institute of Artificial Intelligence Innovation, Zhipu and other partners, Fenghe is described as the world’s first open-source meteorological large language model at the 100-billion-parameter scale.
Unlike conventional numerical weather prediction models, Fenghe is designed more like an “AI meteorological service officer.” Built on a large language model architecture, it combines artificial intelligence with professional meteorological knowledge and massive amounts of weather and climate data to support weather analysis, risk assessment and decision-making.
Wang Muhua, a senior engineer at the CMA Public Meteorological Service Centre, said one of Fenghe’s key features is its foundation model with hundreds of billions of parameters. Through multimodal integration and generative AI technologies, the model is designed to improve the resolution, efficiency and responsiveness of meteorological services.
Fenghe is built on a comprehensive Earth system data infrastructure and has been trained on 50 million tokens of high-quality meteorological service data. It also integrates authoritative meteorological datasets and has completed the required filing process for generative AI services in China, providing users with a more secure and controllable model application environment.
Fenghe differs in its positioning from several AI-based forecasting systems previously developed by the CMA, including Fengqing, a global medium- and short-range forecasting system; Fenglei, an AI-based nowcasting system; and Fengshun, a global subseasonal-to-seasonal prediction system.
These systems are primarily designed for professional meteorological operations and focus on improving forecasting capabilities across different time scales. Fenghe, by contrast, is aimed mainly at the public and industries, serving as an intelligent interface between professional weather forecasts and real-world service needs.
According to Wang, Fenghe is not intended to replace traditional numerical weather prediction, nor does it simply use a large language model to generate weather forecasts. Instead, it builds on forecasting information produced by systems such as Fengqing, Fenglei and Fengshun, while combining professional meteorological knowledge with generative AI and application scenarios. This approach is intended to address limitations of general-purpose large language models, which may struggle to fully understand specialized meteorological needs, generate sufficiently professional information or adapt effectively to complex service scenarios.
For the public, this could mean a shift from simply being told what the weather will be to receiving practical advice on what to do. Yu Tingzhao, a senior engineer at the CMA Public Meteorological Service Centre, said Fenghe could provide more precise and scenario-specific support for travel and outdoor activities by combining weather information with location, timing and individual needs.
For example, a conventional forecast might say that an area will be cloudy with occasional showers over the weekend. For someone planning a hike or camping trip, such information may not be sufficient to determine whether or where to go. In the future, Fenghe could combine high-resolution weather forecasts with local environmental and geographical information to analyze conditions in complex areas such as mountains and lakes. It could identify differences in rainfall, wind speed and visibility between locations and time periods, and then provide practical recommendations, such as choosing a particular route or completing a hike before deteriorating conditions arrive.
This could gradually shift weather services from regional forecasts toward highly localized, point-specific services, and from passive information delivery toward proactive, AI-assisted decision-making.
Fenghe is already supporting meteorological services across China and providing the public with personalized weather information, service recommendations and risk warnings. Its international version has also been launched and integrated into “Mazu,” an intelligent meteorological early-warning solution developed in support of the Early Warnings for All initiative. It provides users around the world with bilingual Chinese-English intelligent question answering, weather information and risk analysis.
With the launch of the global open-source initiative, the CMA plans to make Fenghe’s complete model weights available through platforms including GitHub, Hugging Face and ModelScope. Standardized APIs, cloud services and customized deployment solutions will also be provided. The initiative therefore goes beyond simply opening the model code and weights: it aims to provide a complete technology and deployment framework that allows developers, research institutions and international partners to integrate Fenghe into applications ranging from mobile apps and mini-programs to embodied AI systems.
The broader goal is to build an open and collaborative global ecosystem for meteorological artificial intelligence. By making advanced meteorological AI capabilities more accessible, Fenghe could help transform large volumes of weather and climate data into intuitive risk information, convert complex warnings into easy-to-understand natural-language guidance, and eventually translate weather risks into concrete actions.
In sectors such as transportation, energy, electricity, healthcare, logistics and tourism, meteorological warnings could be further connected with operational decisions, enabling AI systems to recommend specific responses to changing weather conditions. From public weather services to professional analysis, and from risk identification to emergency response, Fenghe is intended to make meteorological warnings more understandable, accessible and actionable, providing new technological support for global disaster risk reduction and the development of more inclusive early-warning systems.
Source: people, xinhua, yicai, ceic, kpzg, cctv
