IBM released Granite Time Series PatchTST-FM-r2. This open foundation model targets enterprise time-series workloads. These include demand forecasting, energy load prediction, and pricing analysis.
The model contains approximately 385 million parameters. It operates in a zero-shot setting, applying to new time-series data without task-specific fine-tuning.
To encourage adoption, IBM made the model's weights, architecture, and code openly available. They are released under permissive Apache 2.0 and OpenMDW 1.0 licenses. This aims to lower the barrier for organizations to use, modify, or distribute the model.
IBM reported strong performance on the GIFT-Eval benchmark. It ranked second among replicable zero-shot models.