IBM Research has released Granite Timeseries Ensemble, an open-source approach that combines four complementary pre-trained time-series models to improve forecasting accuracy without requiring dataset-specific retraining. The ensemble addresses a common problem in time-series machine learning: selecting the best model for a given task, which often involves trial and error since different architectures excel at different patterns—some handle smooth industrial signals better than noisy retail demand, and others perform well across different seasonal horizons. The ensemble combines four Granite architectures—PatchTST-FM-r1, PatchTST-FM-r2, FlowState, and Tiny Time Mixers—each designed to capture trends, seasonality, and local variations in different ways. By pooling their predictions, the system reduces the risk of choosing a poorly-performing single model. IBM offers two combination methods: the Probabilistic Ensemble, which weights all models equally and serves as the recommended default, and an adaptive Inter Quartile Range ensemble that gives narrower forecast intervals greater influence. Both approaches avoid the complexity of fitting weights from historical data. The models are permissively licensed and available on Hugging Face, with a simple Python interface that loads the pre-trained ensemble directly. Users can install the Granite TSFM package and deploy forecasts with just a few lines of code, making professional-grade ensemble forecasting accessible to organizations without deep expertise in model selection or ensemble construction.