Monitoring brain metastases is a time-consuming process, especially when relying on manual analysis of multiple small lesions. While Response Assessment in Neuro-Oncology Brain Metastases (RANO-BM) guidelines recommend unidimensional measurements, volumetric assessment of tumors and surrounding edema is critical for informed clinical decisions. Manual methods, however, are often error-prone and subjective, motivating the development of automated, reproducible artificial intelligence-powered solutions for monitoring the progress of the disease. In this work, we address the problem of automated brain metastasis detection and multi-class segmentation using the nnU-Net framework—a self-configuring deep learning model for medical image analysis. We evaluate different configurations of nnU-Net, including changes to loss functions, model size, training duration, and post-processing routines. Our experiments (GLI-team) on the BraTS-METS 2025 dataset demonstrate robust performance of nnU-Nets across diverse MRI scanners and clinical protocols (exceeding the Dice score of 0.78 for the enhancing part of the tumor). This heterogeneity makes BraTS-METS a crucial benchmark for assessing model generalizability. Our results highlight the promise of automated, objective tools in supporting more consistent and scalable clinical workflows for managing patients with brain metastases. Finally, our approach outperformed all other techniques and was the winning solution in the BraTS-METS 2025 Challenge.