IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license
huggingface ·
IBM has released Granite Time Series PatchTST-FM-r2, a new ~385M-parameter time-series foundation model that achieves strong zero-shot forecasting performance on the GIFT-Eval benchmark. As of September 8, 2026, it ranks as the top zero-shot model under a permissive, commercial-friendly open-source license and second overall among replicable zero-shot models, while also remaining competitive against models allowed to train on benchmark data. The architecture builds on PatchTST-FM-r1 by using conformer-style blocks that combine multi-head self-attention with temporal convolution, overlapping patches with Hamming-window weighting, and support for long contexts up to 8,192 steps with probabilistic forecasts across 99 quantiles. Its pretraining corpus is documented and includes selected datasets, synthetic data from KernelSynth and CauKer, and a TSMixup corpus restricted to datasets outside the GIFT-Eval evaluation set. The model is dual-licensed under Apache 2.0 and OpenMDW 1.0, and its weights, architecture, inference pipeline, and reproduction code are openly available. The article also provides Python code examples for using the model from the Hugging Face Hub and notes that several Granite Time Series models are available through an Early Access program with Confluent for streaming applications via Apache Flink.
huggingface