IBM Releases Granite PatchTST-FM-r2 Time-Series Model Under Apache 2.0 and OpenMDW

IBM published Granite Time Series PatchTST-FM-r2 on September 9, 2026, via a Hugging Face research blog, releasing a roughly 385-million-parameter zero-shot forecasting model dual-licensed under Apache 2.0 and OpenMDW 1.0. Users may pick either license. Weights live at ibm-granite/granite-timeseries-patchtst-fm-r2, and the Hub card tags license:openmdw-1.0.

PatchTST-FM-r2 succeeds PatchTST-FM-r1 with a redesigned backbone, a larger pretraining corpus, probabilistic forecasts, missing-value imputation, and documented training data. IBM positions it for demand, prices, energy, traffic, telemetry, and similar regularly sampled series without per-dataset fine-tuning.

Benchmarks and architecture

As of September 8, 2026, IBM reports PatchTST-FM-r2 as the top-performing zero-shot model with a permissive commercial-friendly open-source license among replicable zero-shot entries on GIFT-Eval, and second overall among replicable zero-shot models, with geometric-mean CRPS 0.467 and MASE 0.6846 on that leaderboard slice. Those rankings are IBM’s reading of GIFT-Eval as of that date.

Architecturally, r2 replaces standard transformer blocks with conformer-style layers that combine multi-head self-attention and temporal convolution, uses 50% overlapping patches with Hamming-window overlap-and-add forecasting, expands from 20 to 30 blocks, supports contexts up to 8,192 steps, and emits 99 quantiles for uncertainty intervals.

Data, code, and streaming

IBM documents a pretraining mix of selected GiftEvalPretrain datasets, KernelSynth-style synthetics, a TSMixup corpus restricted outside GIFT-Eval evaluation sets, and about 500,000 CauKer sequences. Architecture code ships in the Granite-TSFM repository and stays backward-compatible with r1 checkpoints. Install path in the blog is pip install "granite-tsfm>=0.3.9" then PatchTSTFMForPrediction.from_pretrained("ibm-granite/granite-timeseries-patchtst-fm-r2"). Related Granite time-series models also appear in an IBM–Confluent Cloud Early Access for streaming Flink inference; PatchTST-FM-r2 itself is the notebook and Hub release described here.

Primary sources for this brief are IBM’s September 9 Hugging Face blog post and the ibm-granite/granite-timeseries-patchtst-fm-r2 model card.