IBM releases Granite Time Series PatchTST-FM-r2, a 385M-parameter open zero-shot forecasting model ranking #2 on GIFT-Eval

IBM released Granite Time Series PatchTST-FM-r2, a roughly 385-million-parameter open-weight time-series forecasting model, as the successor to PatchTST-FM-r1. It has a redesigned architecture combining self-attention and temporal convolution, handles context lengths up to 8,192, produces probabilistic forecasts across 99 quantiles, and can handle missing data. Weights, code and a reproduction pipeline are published on Hugging Face under dual Apache-2.0/OpenMDW-1.0 licences, and IBM reports it ranks #2 on the GIFT-Eval leaderboard among zero-shot models.

This gives developers and enterprises a single pre-trained model that can forecast on new datasets without retraining, and the permissive commercial licence removes a real barrier that restrictive research licences often impose. It matters most to teams doing demand forecasting, capacity planning or anomaly detection who want to avoid building bespoke per-dataset models. The practical upside is real but narrow: this is one component in a forecasting pipeline, not a general capability shift, and adoption depends on integration effort and how it performs on a given company's actual data rather than benchmark data.

The "SOTA" framing in the headline is true only within a carefully carved-out subset (permissive-licence, replicable, zero-shot models), and the model actually ranks #2 overall, not #1. It's an incremental r1-to-r2 update in an already crowded field (TimesFM-3, Chronos-2, Timer-S1, Toto all compete here), and the Confluent integration mention reads as a partnership plug rather than independent validation. Coverage should be read as a solid technical release, not a breakthrough.

1. 01Whether independent users replicate the claimed GIFT-Eval metrics (CRPS 0.467, MASE 0.6846) outside IBM's own benchmarking.
2. 02Download and usage figures on Hugging Face over the next 1-3 months compared to Chronos-2 and TimesFM-3 to gauge real adoption versus PR interest.
3. 03Whether IBM or third parties publish head-to-head comparisons on real enterprise datasets (not just GIFT-Eval) showing performance against fine-tuned, per-dataset models.