IBM has released the foundation model for time series forecasting, "Granite Time Series PatchTST-FM-r2." This model, featuring approximately 385 million parameters, is an improved version of the existing PatchTST-FM-r1.
Model ReleasesIBMGranite Time Series PatchTST-FM-r2
IBM Releases Commercial-Ready Time Series Foundation Model Granite Time Series PatchTST-FM-r2
This article is a translation. Read the Japanese original
The new model incorporates Conformer layers into its architecture. By combining Self-attention with convolutional layers, the Conformer can efficiently capture both long-term relationships and local temporal structures. Additionally, the model aims to improve forecasting accuracy by utilizing patch representations with a 50% overlap and weighting via a window function (Hamming-window).
In the "GIFT-Eval" benchmark, this model recorded the highest performance among replicable, zero-shot models with commercially usable licenses. The model weights, architecture, inference pipeline, and code to reproduce the benchmark results have all been made public.
Sources
- IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license （Hugging Face Blog, 2026-09-10）
- IBM Granite GitHub