IBM has released Granite Time Series PatchTST-FM-r2, a ~385M-parameter time series foundation model that achieves state-of-the-art zero-shot performance on the GIFT-Eval leaderboard under permissive commercial licenses.
- The model uses an updated architecture with conformer blocks combining multi-head self-attention and temporal convolution to capture long- and short-range relationships.
- It supports contexts up to 8,192 steps and generates probabilistic forecasts via a 99-quantile prediction head.
- PatchTST-FM-r2 ranks #2 overall among replicable zero-shot models on GIFT-Eval and is the top performer among those with permissive licensing.
- The pretraining corpus includes synthetic data and datasets outside the GIFT-Eval evaluation set to prevent benchmark leakage.
- Weights, architecture, and inference code are available under dual Apache 2.0 and OpenMDW 1.0 licenses.
The release aims to reduce adoption barriers for enterprises by providing transparent training data documentation and commercial-friendly licensing alongside streaming integration capabilities.