It is loaded through granite-tsfm>=0.3.11 as ibm-granite/granite-timeseries-ensemble-r1; the configuration downloads four checkpoints at runtime: Granite PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1, and TTM-r3.
The default Probabilistic Ensemble pools member quantile forecasts with equal influence and requires neither retraining nor fitted weights. An optional IQR Ensemble changes influence by forecast spread. Both return point forecasts and quantiles, with the default output spanning 0.1 to 0.9. The recipe and its default Granite members are described as permissively licensed; expanded recipes can carry separate non-commercial or research-use terms.
On 97 GIFT-Eval configurations, the default ensemble reported MASE of 0.683, versus 0.685 for PatchTST-FM-r2, the strongest individual member on that measure. PatchTST-FM-r2 reported the best CRPS/MWQL, 0.467 versus 0.471 for each ensemble variant. These aggregate results do not establish accuracy, calibration, latency, memory use, or licensing suitability for a particular dataset or deployment.