Granite Time Series PatchTST-FM R2
PatchTST-FM R2 is a roughly 385M zero-shot forecasting model. It builds on IBM's earlier PatchTST foundation model and adds Conformer blocks that combine attention with local convolution.
Model design and output
The model has 30 Conformer blocks and a hidden dimension of 1,024. It uses patches of length 16 with stride 8. Overlapping predictions are combined through an overlap-add method. The model card permits users to choose either OpenMDW 1.0 or Apache 2.0.Granite Time Series PatchTST-FM R2 model card
Practical use
Practical guidance. Use quantiles to examine forecast uncertainty as well as central estimates. Check interval coverage on held-out periods, especially around shifts and unusual events. Its larger parameter count makes throughput and batching worth measuring alongside forecast quality.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ibm-granite/granite-timeseries-patchtst-fm-r2 |
| License | Choice of OpenMDW 1.0 or Apache 2.0 under the model card's dual-license terms.Granite Time Series PatchTST-FM R2 model card |
| Training context | 8,192 time-series observationsGranite Time Series PatchTST-FM R2 model card |
| Probabilistic output | 99 quantilesGranite Time Series PatchTST-FM R2 model card |
Related models
- Granite Time Series FlowState R1.1
IBM · Granite Time Series
An 18.5M continuous forecasting model with a state-space encoder and functional-basis decoder.
Launch date not confirmedDownloadable - Granite Time Series TTM R3
IBM · Granite Time Series
A forecasting system that routes numerical time series to compact TinyTimeMixer experts.
Launch date not confirmedDownloadable