IBM releases Granite TimeSeries PatchTST-FM-r2

IBM has released Granite TimeSeries PatchTST-FM-r2, a roughly 385-million-parameter foundation model for zero-shot time-series forecasting. The previous r1 model was listed at about 260 million parameters.

R2 changes the model’s architecture and training data but keeps several core settings from r1. Both versions use an 8,192-step context window, a hidden dimension of 1,024, patches of length 16 and a 99-quantile output head.

IBM expanded the stack from 20 blocks to 30 and replaced r1’s plain Transformer blocks with Conformer blocks. According to IBM’s model card, each block places a convolution layer after multi-head self-attention and between two half-step feed-forward networks, with convolution kernels in a repeating 5, 5, 3, 3 pattern. R2 also adds pre-head layer normalization, which IBM says improves training stability.

IBM says r2 uses overlapping patches with length 16 and stride 8, Hamming-window loss weighting during training, and an overlap-and-add method to combine forecasts at inference. The company also added a fourth training-data source: an internal CauKer dataset that it describes as about 500,000 synthetic sequences, each 4,096 steps long. IBM says the implementation remains compatible with r1 checkpoints, but that claim has not been independently tested.

The model card says users may choose between OpenMDW License 1.0 and Apache License 2.0. IBM calls the licensing permissive and commercial-friendly; that is the issuer’s characterization, not an independent legal conclusion about suitability for a particular commercial use.

IBM also says r2 ranked second among replicable zero-shot models on both CRPS and MASE in GIFT-Eval as of Aug. 31, 2026. IBM’s model card says the results were in a pending pull request. The current GIFT-Eval repository now lists an r2 results entry, but the exact historical rankings and IBM’s comparisons with pretrained models were not independently reproduced.