IBM released Granite Time Series PatchTST-FM-r2, a ~385M-parameter time-series foundation model dual-licensed under Apache 2.0 and OpenMDW 1.0. As of September 8, 2026, it ranked second among replicable zero-shot models on the GIFT-Eval benchmark and first among permissively licensed models in that category. The model replaces PatchTST-FM-r1's transformer blocks with conformer layers combining self-attention and temporal convolution, and adds probabilistic forecasting and missing-value imputation. Weights, code, and benchmark reproduction materials are publicly available.
No score is assigned. Sources and their independence are shown in the citation chain below.