IBM Releases Groundbreaking Time-Series Forecasting Model PatchTST-FM-R2

IBM has released its latest time-series forecasting model, Granite PatchTST-FM-R2. The zero-shot model is available under a dual license from Apache 2.0 and OpenMDW 1.0. It combines an updated architecture, a larger pretraining corpus, probabilistic forecasting, and support for imputation of missing values.

The model uses conformer blocks to capture short-range temporal structure, allowing the attention mechanism to focus on long-range relationships between patches. According to the developers, PatchTST-FM-R2 achieves state-of-the-art results on GIFT-Eval, a comprehensive benchmark for evaluating forecasting models. As of September 8, 2026, it ranks second among replicable zero-shot models and third when including pretrained models.

The model retains 50% overlapping patches with Hamming-window weighting during training and overlap-and-add forecasting at inference. It has a hidden dimension of 1024 and 30 blocks, supports context lengths up to 8,192 steps, and predicts 99 quantiles over flexible forecast lengths.