IBM's Granite PatchTST-FM-r2 tops commercially licensed zero-shot forecasters on GIFT-Eval
IBM released Granite Time Series PatchTST-FM-r2, a ~385M-parameter forecasting foundation model dual-licensed under Apache 2.0 and OpenMDW that ranks second among replicable zero-shot models on GIFT-Eval and first among those with commercial-friendly licenses. The edge is licensing, not raw accuracy — it trails only models with more restrictive terms, and ships open weights plus a documented pretraining corpus that enterprise governance reviews require.
Source: huggingface.co ↗
PatchTST-FM-r2 is the highest-performing model in the same category among models with permissive, commercial-friendly licensing.
IBM Granite blog
Why this matters
- → Enterprise teams can now use a SOTA forecaster without restrictive licensing overhead.
- → Documented pretraining corpus lets governance teams audit what data shaped the model.
- → Conformer architecture (attention + convolution) outperforms pure transformers on time series.
Open forecasting for enterprise