IBM has released Granite Time Series PatchTST-FM-r2, a 385-million-parameter model built to forecast unfamiliar time series without task-specific training. It supports contexts up to 8,192 steps, flexible forecast lengths, missing-value imputation and probabilistic output across 99 quantiles.
The model combines attention for long-range relationships with temporal convolution for short-range patterns. Overlapping patches and weighted reconstruction are intended to smooth boundaries between predicted segments. IBM says the system ranks second among replicable zero-shot models on two GIFT-Eval measures and first within that category among models carrying permissive, commercially friendly licenses.
The comparison is benchmark-specific rather than proof that one model is best for every demand, price, energy or telemetry dataset. IBM has published the weights, architecture, inference pipeline and code needed to reproduce the results, allowing teams to test that claim on their own data. Users may choose Apache 2.0 or OpenMDW 1.0 licensing. The release lowers the barrier to evaluating a general forecasting model before maintaining separate models for every dataset.