IBM Releases PatchTST-FM-r2 Time Series Model

IBM released Granite Time Series PatchTST-FM-r2, a new foundation model for zero-shot forecasting. The model succeeds PatchTST-FM-r1 and is available under the Apache 2.0 and OpenMDW 1.0 licenses.

As of September 8, 2026, the ~385M-parameter model ranks #2 overall among replicable zero-shot models on the GIFT-Eval benchmark. It achieves a geometric-mean CRPS of 0.467 and a geometric-mean MASE of 0.6846, making it the highest-performing model in its category with a permissive open-source license. When evaluated alongside models permitted to train on benchmark data, it ranks 3rd for CRPS and 4th for MASE.

PatchTST-FM-r2 replaces standard transformer layers with conformer-style blocks that combine multi-head self-attention with temporal convolution layers. The backbone uses alternating convolution kernel sizes of 3 and 5 in a repeating pattern of 5, 5, 3, 3, backed by an expansion from 20 to 30 blocks. The architecture includes 50% overlapping patches with Hamming-window weighting, overlap-and-add forecasting, missing value imputation, and support for contexts up to 8,192 steps while predicting 99 quantiles.

Pretraining data includes selected datasets from GiftEvalPretrain, KernelSynth synthetic data, a TSMixup corpus, and approximately 500,000 synthetic CauKer sequences of length 4,096.

Read the original announcement