IBM releases Granite Time Series PatchTST-FM-r2 model with Apache 2.0 and OpenMDW 1.0 licensing

IBM has released Granite Time Series PatchTST-FM-r2, the latest model in the Granite Time Series Foundation Model (TSFM) family, on Hugging Face. The model is a ~385M-parameter time-series foundation model designed for zero-shot forecasting across diverse domains such as demand, prices, energy loads, traffic, and telemetry.

PatchTST-FM-r2 introduces a redesigned architecture using conformer blocks, which combine multi-head self-attention with temporal convolution to capture both long- and short-range temporal structure. The model also expands from 20 to 30 blocks and uses 50% overlapping patches with Hamming-window weighting and overlap-and-add forecasting to smooth patch boundaries.

As of September 8, 2026, PatchTST-FM-r2 ranks second overall for CRPS and fourth for MASE among replicable models on the GIFT-Eval benchmark when including both zero-shot and pretrained models. Among zero-shot, replicable models with permissive licenses, it ranks first for both metrics.

The model is dual-licensed under Apache 2.0 and OpenMDW 1.0, both of which are permissive and commercially friendly. IBM provides open weights, architecture, inference pipeline, and code to reproduce benchmark results.

The release highlights the model's strong zero-shot performance on GIFT-Eval, a comprehensive time-series forecasting benchmark, and includes code examples for usage in notebooks and streaming applications.