IBM released Granite Time Series PatchTST-FM-r2 on Sept. 9, 2026, as an open zero-shot forecasting model in its Granite Time Series family. The company published weights, architecture, an inference pipeline, and code it says can reproduce the cited benchmarks.
PatchTST-FM-r2 is a time-series foundation model, not a chat LLM. It takes a recent history of regularly sampled numbers such as demand, energy load, prices, traffic, or telemetry and returns future values plus uncertainty ranges, without training a separate model for every dataset. Unlike chat models that generate free-form text, it forecasts numerical sequences and can fill missing points in the input history.
IBM says the model has about 385 million parameters, supports contexts up to 8,192 steps, flexible forecast lengths, probabilistic forecasts through a 99-quantile prediction head, and missing-value imputation. In r2, IBM says it replaced standard transformer layers with 30 conformer-style blocks that combine multi-head self-attention and temporal convolution, and uses 50% overlapping patches with Hamming-window weighting and overlap-and-add forecasting.
As of Sept. 8, 2026, IBM reports PatchTST-FM-r2 ranked second for both CRPS and MASE among replicable zero-shot models evaluated without test leakage on GIFT-Eval, and first in that slice among models with permissive commercial-friendly licensing. IBM reports geometric-mean scores of 0.467 CRPS and 0.6846 MASE, where lower is better. That ranking is IBM-reported for the stated slice and date; it is not a claim that the model is best for every forecasting task.
An independent MustHave.ai review dated Sept. 13, 2026, repeats the core specifications and advises treating the GIFT-Eval ranking as IBM-reported because the model card said the result was in a pending benchmark pull request. MustHave.ai said it did not independently rerun the evaluation. IBM says the model is dual-licensed under Apache 2.0 and OpenMDW 1.0; the release still notes that organizational licensing and model-governance review remains necessary.
What remains open is when GIFT-Eval maintainers will merge and display the pending result, how well the 99-quantile intervals calibrate on a given organization's own data, and whether zero-shot forecasts beat a simple seasonal baseline under that organization's missing-data and regime-shift patterns. IBM's leaderboard scores and MustHave.ai's coverage should be weighed separately until those points are settled.