IBM Releases Granite PatchTST-FM-R2 Zero-Shot Time Series Model

IBM released Granite Time Series PatchTST-FM-r2 on September 9, 2026, a roughly 385-million-parameter zero-shot time-series forecasting model dual-licensed under Apache 2.0 and the Linux Foundation’s OpenMDW 1.0. The model weights, architecture, inference pipeline, and code needed to reproduce its benchmark results were all published openly with the release.

IBM announced the model in a Hugging Face blog post by IBM Research authors, presenting it as the successor to PatchTST-FM-r1 and the latest model in the Granite time-series foundation model family. According to the announcement, PatchTST-FM-r2 combines an updated architecture, a larger pretraining corpus, probabilistic forecasting, and support for imputation of missing values, and it generates forecasts on new data without fine-tuning or task-specific fitting.

Reported GIFT-Eval Standings

GIFT-Eval is a comprehensive benchmark for evaluating forecasting models across diverse datasets and scenarios, and lower values are better for both CRPS and MASE, the two metrics IBM reports. IBM said that as of September 8, 2026, PatchTST-FM-r2 ranks second among replicable, zero-shot models for both metrics, and is the highest-performing model in that category among models released under permissive, commercial-friendly licenses. The announcement reports a geometric-mean CRPS of 0.467, immediately behind TimesFM-3, and a geometric-mean MASE of 0.6846.

Some models on GIFT-Eval are categorized as pretrained rather than strictly zero-shot, meaning they are allowed to include the training portions of the benchmark’s evaluation datasets in their pretraining corpora. IBM said that even when those models are added to the comparison, PatchTST-FM-r2 places third for CRPS and fourth for MASE among replicable models, ahead of pretrained Chronos-2, Timer-S1, and Toto variants, some of them considerably larger.

The model card, authored by Jiri Navratil, Wesley M. Gifford, Yunshi Wen, Chandra K. Reddy, and Agung Julius, dates the number-two replicable zero-shot standing to August 31, 2026, and notes that the model’s results sit in a pending pull request submitted to the GIFT-Eval benchmark; the announcement dates the same standing to September 8, 2026.

Conformer Architecture

PatchTST-FM-r2 retains the patch-based representation of its predecessor but replaces standard transformer blocks with conformer blocks, a design that originated in speech processing. Each block contains two half-step feed-forward layers surrounding multi-head self-attention and a temporal convolution layer, with alternating convolution kernel sizes of 3 and 5 in a repeating {5, 5, 3, 3} pattern. IBM said the convolution component captures short-range temporal structure, allowing the attention mechanism to concentrate on long-range relationships between patches.

The model uses 50% overlapping patches — patch length 16, stride 8 — with Hamming-window weighting during training and overlap-and-add forecasting at inference, plus a pre-head layer norm applied for training stability. It has a hidden dimension of 1024 and 30 blocks, up from 20 in r1, supports context lengths up to 8,192 steps, and predicts 99 quantiles over flexible forecast lengths, producing both point forecasts and uncertainty intervals. The implementation is available in the open-source granite-tsfm repository and remains backward compatible with PatchTST-FM-r1 checkpoints.

Training Data and Licensing

The documented pretraining corpus has four sources: selected datasets from GiftEvalPretrain; custom synthetic data based on KernelSynth with modified periodic kernels and limited augmentation; a TSMixup corpus generated using the approach described in the Chronos paper but restricted to datasets outside the GIFT-Eval evaluation set; and approximately 500,000 synthetic CauKer sequences, each with a length of 4,096. The model card notes that the CauKer dataset was generated by IBM’s FlowState team, and it cites the 2026 arXiv paper “Revisiting the Generic Transformer: Deconstructing a Strong Baseline for Time Series Foundation Models” for the underlying approach.

Users may choose either the Apache 2.0 license or OpenMDW 1.0, a Linux Foundation licensing framework designed for AI models and related materials. IBM said both licenses grant broad, permissive rights to use, modify, and distribute the model, and that it aims to reduce barriers to adoption. A disclosure on the model card states that IBM developers produced the code as an open-source project rather than an IBM product, and that IBM is under no obligation to provide enhancements, updates, or support.

Availability and Granite Family Context

The weights are downloadable from the Hugging Face Hub and load through the granite-tsfm package, version 0.3.9 or later, via a pipeline API. IBM’s example code generates a forecast, including requested quantiles, from the last 512 samples of an ETTh1 series, and IBM positions the model for demand, prices, energy loads, traffic, telemetry, and other regularly sampled time series.

For streaming deployments, IBM and Confluent have made several Granite Time Series models (PatchTST-FM-r1, FlowState-r1.1, TTM-r3, and TSPulse) available through an Early Access program in Confluent Cloud, which runs foundation-model inference on live streams through Apache Flink.

An IBM Research overview of the family documents the lineage behind the release: TST in 2021, among the first transformer-based models applied to time-series data; PatchTST in 2023, which introduced patching and channel independence; Tiny Time Mixer in 2024; and FlowState in 2025. According to that overview, the models have collectively been trained on more than 100 billion data points, and TTM models have passed 37 million downloads on Hugging Face. PatchTST-FM-r1, the new model’s direct predecessor, was co-developed with Rensselaer Polytechnic Institute, and IBM’s research-version models carry a non-commercial license while the Granite line is published under Apache 2.0.