IBM releases Granite Time Series PatchTST-FM-r2 at about 385M parameters

IBM Research published Granite Time Series PatchTST-FM-r2 — about 385 million parameters — on a Hugging Face blog and model card, with GIFT-Eval ranks and “SOTA” lines staying IBM’s dated claims.

An IBM weight drop plus blog is a product filing you can download and inspect.

IBM Research published “IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license” on the Hugging Face blog, with weights at ibm-granite/granite-timeseries-patchtst-fm-r2. Hugging Face is a public model-hosting site. Time series means numbers over time — sales, sensors, weather. That pair is the primary. SOTA means “state of the art,” which is IBM’s headline.

IBM says the model is about 385M parameters — 385 million tunable weights inside the network. Specs on the card and blog: context up to 8,192 time steps; 30 conformer blocks (up from 20 in the r1 release); 50% overlapping patches, so neighboring windows of the series share half their data; a 99-quantile prediction head, which estimates a full spread of outcomes rather than a single point; and a dual license, Apache 2.0 or OpenMDW 1.0. Install pins and library versions are in Sources.

GIFT-Eval is a public forecasting benchmark. Those lines are IBM/Hugging Face’s, and the dates differ across the two primaries. The model card: as of 31 August 2026, top-performing zero-shot model under a permissive commercial-friendly license, and #2 among replicable zero-shot models on CRPS and MASE (two error scores — lower is better). The blog: as of 8 September 2026, the same #2 / first-among-permissive framing, with geometric-mean CRPS 0.467 and MASE 0.6846. IBM also says that with pretrained models included it ranks 3rd on CRPS and 4th on MASE among replicable entries. The card notes the r2 scores were in a pending GIFT-Eval pull request at publication.