Granite Time Series TTM R3
Granite TTM R3 is a forecasting release built from compact TinyTimeMixer models. A lightweight router selects or blends experts based on the input series. The full models range from 1M to 35M parameters.
Model design and output
The family uses linear and mixing components rather than full self-attention. It supports zero-shot forecasting, few-shot adaptation and full fine-tuning. Multivariate inputs and external variables can inform forecasts. Parameter ranges describe the expert collection, not one fixed checkpoint.Granite Time Series TTM R3 model card
Practical use
Practical guidance. Select a forecast horizon and data frequency before comparing experts. Use a chronological holdout so future observations cannot leak into the input. Assess forecast error against a simple seasonal baseline, then check whether external variables are actually known at prediction time.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ibm-granite/granite-timeseries-ttm-r3 |
| License | Apache 2.0Granite Time Series TTM R3 model card |
| Lite models | 1M to 18M parametersGranite Time Series TTM R3 model card |
| Input and output | Numerical time-series history and forecastsGranite Time Series TTM R3 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 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