IBM Publishes Granite Time Series PatchTST-FM-r2 Under a Commercial-Friendly License
IBM's new ~385M-parameter time-series foundation model ranks second among replicable zero-shot models on GIFT-Eval and ships under Apache 2.0.
IBM has published Granite Time Series PatchTST-FM-r2, the newest member of its Granite time-series family, on Hugging Face. The model is distributed under Apache 2.0 and OpenMDW 1.0, making it available for commercial use. Compared with its predecessor PatchTST-FM-r1, it brings a redesigned architecture, a larger pretraining corpus, probabilistic forecasting and support for imputation of missing values.
The model has roughly 385 million parameters and supports contexts of up to 8,192 steps. The idea behind time-series foundation models is to replace one trained model per dataset with a single pretrained model that produces zero-shot forecasts.
GIFT-Eval results
According to IBM, as of September 8, 2026 the model ranks second on both CRPS and MASE among replicable, zero-shot models on the GIFT-Eval leaderboard, with a geometric-mean CRPS of 0.467 and MASE of 0.6846. It sits immediately behind TimesFM-3 in that comparison and comes first among models with permissive licensing.
When models classified as pretrained, which may include the training splits of the evaluation datasets in their corpora, are added to the comparison, PatchTST-FM-r2 places third for CRPS and fourth for MASE. IBM says it outperforms Chronos-2, Timer-S1 and Toto variants, some of which are considerably larger.
What changed in the architecture
The patch-based representation of the PatchTST family remains, but standard transformer layers were replaced with conformer layers borrowed from speech processing. Each block wraps multi-head self-attention and a temporal convolution layer between two half-step feed-forward layers. Convolution handles local temporal structure while attention concentrates on longer horizons. Kernel sizes alternate in a repeating {5, 5, 3, 3} pattern, and the backbone grew from 20 to 30 blocks.
The model also uses 50% overlapping patches with Hamming-window weighting and overlap-and-add forecasting to smooth patch boundaries. Weights, the inference pipeline and the code needed to reproduce the benchmark numbers are all available. Full details are in the Hugging Face blog post.
