IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license

IBM has introduced Granite Time Series PatchTST-FM-r2, a 385M-parameter time-series forecasting model designed for zero-shot forecasting across diverse datasets. The model builds on its predecessor by incorporating probabilistic forecasting, missing value imputation, and an updated architecture that combines convolution with multi-head self-attention. These changes improve the model’s ability to capture both short- and long-term relationships in time series data, enhancing forecasting accuracy.

PatchTST-FM-r2 ranks second overall among replicable, zero-shot models on the GIFT-Eval leaderboard as of September 8, 2026, with a geometric-mean CRPS of 0.467 and MASE of 0.6846. It outperforms several larger pretrained models despite being zero-shot, demonstrating strong generalization across heterogeneous datasets. The model’s permissive Apache 2.0 and OpenMDW 1.0 licenses allow for both research and commercial use without restrictions.

The model’s architecture replaces standard transformer layers with conformer layers, which integrate convolution and multi-head self-attention to better model local and global temporal patterns. It uses 50% overlapping patches with Hamming-window weighting and overlap-and-add forecasting to smooth patch boundaries, improving forecast accuracy. The model supports contexts up to 8,192 steps and predicts 99 quantiles for flexible forecast lengths, providing both point forecasts and uncertainty intervals.

IBM provides the model weights, architecture, inference pipeline, and code to reproduce benchmark results, along with examples for notebooks and streaming applications. The pretraining corpus includes datasets from GiftEvalPretrain, custom synthetic data, a TSMixup corpus, and approximately 500,000 synthetic CauKer sequences, ensuring no overlap with GIFT-Eval evaluation sets.