IBM has published Granite Time Series PatchTST-FM-r2, a time-series foundation model designed for zero-shot forecasting across domains such as demand, energy load, traffic and telemetry. The company describes the model as containing roughly 385 million parameters, supporting context lengths up to 8,192 steps and producing probabilistic forecasts via a 99-quantile output head.
Performance on GIFT-Eval
IBM reports that PatchTST-FM-r2 ranks second among replicable zero-shot models on the GIFT-Eval benchmark as of September 8, 2026, and is the highest-ranked model in that category released under permissive, commercial-friendly licensing. The model achieves a geometric-mean CRPS of 0.467 and a geometric-mean MASE of 0.6846, according to the published results. When pretrained models that may include benchmark training data are included in the comparison, the model places third for CRPS and fourth for MASE among replicable entries, the company notes.
Architecture and capabilities
PatchTST-FM-r2 builds on the PatchTST family but replaces standard transformer blocks with conformer-style layers that combine multi-head self-attention and temporal convolution. IBM says this design aims to capture both long-range relationships and local temporal structure. The backbone uses alternating convolution kernel sizes in the pattern {5, 5, 3, 3}, expands from 20 to 30 blocks, and applies 50% overlapping patches with Hamming-window weighting and overlap-and-add forecasting to smooth patch boundaries.
The model supports flexible forecast lengths, provides point and quantile outputs for uncertainty estimation, and includes support for probabilistic forecasting and imputation of missing values. IBM indicates the implementation remains backward-compatible with PatchTST-FM-r1 checkpoints.
Training data, licensing and availability
IBM describes the pretraining corpus as four components: selected datasets from GiftEvalPretrain; custom synthetic data from KernelSynth with modified periodic kernels; a TSMixup corpus generated outside the GIFT-Eval evaluation set; and about 500,000 synthetic CauKer sequences of length 4,096. The company emphasizes this documentation to help users assess potential dataset overlap and governance implications.
PatchTST-FM-r2 is published under a dual license (Apache-2.0 and OpenMDW-1.0). IBM states that model weights, architecture, the inference pipeline and code to reproduce benchmark results are available, and that some Granite Time Series models have been integrated for early access streaming inference with Confluent Cloud and Apache Flink. The release date is listed as September 9, 2026.
Original source: Hugging Face Blog