IBM releases Granite PatchTST-FM-r2 time-series forecasting model
IBM has released Granite Time Series PatchTST-FM-r2, a foundation model for zero-shot forecasting of demand, prices, energy loads, traffic, telemetry and other time series. The model has approximately 385 million parameters, supports contexts of up to 8,192 steps and produces forecasts for flexible horizons, including probabilistic outputs through a 99-quantile prediction head. It can also impute missing values.
PatchTST-FM-r2 uses conformer blocks that combine multi-head self-attention with temporal convolution. Compared with PatchTST-FM-r1, it adds 50% overlapping patches, Hamming-window weighting, overlap-and-add forecasting, normalization and 30 backbone blocks. IBM says the model ranks second for both CRPS and MASE among replicable, zero-shot models without test leakage on the GIFT-Eval leaderboard, as of September 8, 2026. It ranks first in that group among models with permissive commercial-friendly licenses.
The model weights, architecture, inference pipeline and reproduction code are available. It is dual-licensed under Apache 2.0 and OpenMDW 1.0. The documented training corpus combines selected GIFT-Eval pretraining data with several synthetic or separately generated datasets.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.