Existing fingerprint-based localization methods perform well in a specific region. However, transferring the model to a new region or adapting to differences in regional environments poses challenges. Additionally, the substantial training cost of the model, including long training time and the inability to reuse data across different regions, further complicates the implementation process. To address these issues, we propose cross region fusion and fast adaptation (CRFA) framework, a novel approach for fingerprint localization in cell-free massive multiple-input multiple-output systems. We begin by extracting angle domain channel power as fingerprint. Further more, we employ access point selection, cross-region fusion and network localization network to enhance localization accuracy and address cross-regional fingerprint disparities. Through the training process of model-agnostic meta-learning, CRFA acquires meta-parameters that facilitate its deployment to any region through a fine-tuning process. Leveraging cross region fusion and meta-learning, the proposed model achieves higher localization accuracy, fast deployment, and adaptability to various environments. Experimental validation using Wireless Insite software shows that the proposed CRFA method performs better in complex environments compared to traditional methods when rapidly deploying models to indoor, urban and suburban region.