IBM's Granite PatchTST-FM-r2 tops zero-shot forecasting leaderboard under Apache 2.0

IBM released Granite Time Series PatchTST-FM-r2, the latest model in its Granite TSFM family, on September 8, 2026. As of that date, the roughly 385-million-parameter model is the top-performing zero-shot model released under a permissive, commercial-friendly open-source license among replicable, zero-shot models on the GIFT-Eval leaderboard, and it ranks second overall among replicable, zero-shot models on that benchmark. The weights, architecture, inference pipeline, and code needed to reproduce the benchmark results are all publicly available.

The release is the successor to PatchTST-FM-r1 and combines an updated architecture, a larger pretraining corpus, probabilistic forecasting, and support for imputing missing values. The model backbone is built from conformer blocks that pair multi-head self-attention with temporal convolution, a design IBM says captures both long- and short-range temporal structure. It supports a context length up to 8,192, flexible forecast lengths, and produces probabilistic forecasts through a 99-quantile prediction head.

The model is dual-licensed under Apache 2.0 and OpenMDW 1.0, and users may select either license. IBM positions the permissive terms as the differentiator: PatchTST-FM-r2 is competitive even against models that are allowed to train on the GIFT-Eval benchmark data, while itself being evaluated strictly zero-shot. The company frames the release against the broader shift in forecasting systems, where users apply a pretrained model to generate forecasts without training and maintaining a separate model for every dataset.

IBM says the model targets general-purpose zero-shot forecasting across demand, prices, energy loads, traffic, telemetry, and other time series. The blog includes code examples for running PatchTST-FM-r2 in a few lines of Python and notes that the Granite Time Series family can be used in streaming production settings through Confluent's product.

The release continues IBM's pattern of iterating on open time-series foundation models. The company points to the model's GitHub repository and a separate blog for the Granite TSFM family, and the model card is available on Hugging Face alongside the announcement. The authors of the blog post are Roman Vaculin, Wesley M. Gifford, Jiri Navratil, Chandra Reddy, and Ayhan Sebin of IBM Research.

What remains to be seen is how the model holds up on the GIFT-Eval leaderboard as other vendors ship competing zero-shot models, and how widely the Apache 2.0 and OpenMDW 1.0 dual license is adopted by enterprises that have historically been cautious about model licensing terms.