Granite Time Series FlowState R1.1
FlowState R1.1 combines a state-space encoder with a functional-basis decoder for continuous forecasting. It has 18.5M parameters. The repository also contains R1.0, which remains the default when no revision is selected.
Model design and output
R1.1 adds CauKer-derived synthetic training data, output gating in its S5 layer and a larger MLP. The decoder represents forecasts through continuous basis functions. This Granite release uses Apache 2.0; the separate IBM Research release has different use terms.Granite Time Series FlowState R1.1 model card
Practical use
Practical guidance. Pin the r1.1 revision explicitly so a deployment does not silently load the earlier model. Compare forecasts at the frequencies and horizons your application needs. Numerical context lengths count observations, and should not be compared with language-model token windows.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ibm-granite/granite-timeseries-flowstate-r1 |
| License | Apache 2.0Granite Time Series FlowState R1.1 model card |
| Required revision | Load revision r1.1 for the updated model.Granite Time Series FlowState R1.1 model card |
| Training context | 4,096 observations; R1.0 used 2,048.Granite Time Series FlowState R1.1 model card |
Related models
- Granite Time Series PatchTST-FM R2
IBM · Granite Time Series
An approximately 385M zero-shot forecasting model that combines attention and local convolution through Conformer blocks.
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