Abstract: Recent advances in generative artificial intelligence and large language models have thoroughly revolutionized human-computer interaction modes and greatly empowered the digital and intelligent transformation of various industrial fields. Although transformer-based general large models deliver strong capabilities in universal language comprehension and open-domain generation, they also display prominent deficiencies when applied in specialized vertical industry scenarios. Due to the lack of systematic accumulation of domain knowledge and targeted scenario adaptation capabilities, these general models frequently produce imprecise demand interpretation, unprofessional service outputs and serious hallucination problems, thereby fail to satisfy the high-precision, standardized and credible service requirements of industrial business scenarios. Accordingly, the research and deployment of domain-specific large models have become an indispensable trend for the in-depth engineering application of generative artificial intelligence technology. Compared with conventional artificial intelligence technologies that rely on manual feature extraction and rule setting, large language models exhibit superior generalization capability, end-to-end learning performance and flexible adaptability to diverse scenarios. They can efficiently uncover implicit logical relationships from massive unstructured data, achieve intelligent understanding of complex semantic information, and substantially reduce the cost of artificial intelligence algorithm development and scenario deployment. These unique technical advantages empower large language models the core driving force for the intelligent upgrading of various vertical industries. Meteorological services are highly suitable for domain-oriented large model deployment. Nevertheless, general-purpose models struggle with meteorological dynamic reasoning, scenario adaptation and domain-specific logical constraints, which hinders the intelligent transformation of meteorological services. To bridge these gaps, Fenghe Meteorological Artificial Intelligence Service System is introduced, which adopts a hundred-billion-parameter mixture-of-experts large language model as the core engine. The system is trained on a high-quality structured meteorological corpus comprising 50 million tokens and 490000 scenario instruction samples. A three-stage training strategy covering domain adaptation, professional fine-tuning and scenario reinforcement learning effectively embeds expert meteorological reasoning and decision-making logic into the model, endowing it with expert-level domain cognitive capabilities. Based upon these core methods, intelligent agents are developed for six key scenarios: Tourism, transportation, energy, logistics, public health, and disaster early warning. Quantitative evaluations verify that Fenghe outperforms mainstream general models by 15%-28% in core professional metrics. The system has been deployed in meteorological departments in more than 20 provinces, supporting major event meteorological guarantees, personalized public services and intelligent industrial decision-making. This work establishes a complete and feasible technical framework for meteorological intelligent services, providing a replicable and scalable technical paradigm for the design, optimization and ecosystem construction of various vertical-domain artificial intelligence systems.