A foundation model for sequences, not ordinary tables
IBM Research released Granite Time Series PatchTST‑FM‑r2 on September 9. The roughly 385M-parameter model supports contexts up to 8,192 time steps, flexible forecast lengths, probabilistic forecasts through a 99-quantile head and missing-value imputation. It targets regularly sampled demand, prices, power loads, traffic and system telemetry without task-specific fine-tuning.
IBM publishes the weights, architecture, inference pipeline and code needed to reproduce its benchmark. Users may select Apache 2.0 or OpenMDW 1.0 under a dual-license arrangement that IBM describes as commercially friendly. Deployers should still read the chosen license and audit data and dependency terms.
“Best” applies to a filtered comparison
IBM reports that, as of September 8, the model ranks second overall among replicable zero-shot models on GIFT‑Eval and first among those with permissive commercial licenses. It lists geometric-mean CRPS of 0.467 and MASE of 0.6846. This is the releasing team's analysis over a defined benchmark filter—not evidence that every business series, regime shift or rare event will be forecast reliably.
Teams should rerun evaluation on their own temporal splits, compare simple seasonal baselines, test leakage and calibrate prediction intervals. Open weights also do not remove inference cost or operational responsibility.