IBM Releases Granite Time Series PatchTST-FM-r2 for Zero-Shot Forecasting

Published September 9, 2026

Overview

Time-series foundation models allow users to apply a pretrained model to new datasets and generate forecasts without training a separate model for each series. IBM has released Granite Time Series PatchTST-FM-r2, the latest addition to the Granite Time Series foundation-model family.

The approximately 385-million-parameter model builds on PatchTST-FM-r1 with an updated architecture, a larger pretraining corpus, probabilistic forecasting, missing-value imputation, and zero-shot forecasting capabilities.

As of September 8, 2026, PatchTST-FM-r2 was the highest-performing zero-shot model released under a permissive, commercial-friendly license among replicable zero-shot models on the GIFT-Eval leaderboard. It ranked second overall in that category. The model is dual-licensed under Apache 2.0 and OpenMDW 1.0.

The model weights, architecture, inference pipeline, and code required to reproduce the benchmark results are available.

Key capabilities

General-purpose zero-shot forecasting for demand, prices, energy loads, traffic, telemetry, and other time series
Approximately 385 million parameters
Context lengths of up to 8,192 steps
Flexible forecast lengths
Probabilistic predictions through a 99-quantile prediction head
Support for missing-value imputation
A conformer-based backbone combining multi-head self-attention with temporal convolution
Open model weights, architecture, inference pipeline, and benchmark-reproduction code
Dual licensing under Apache-2.0 and OpenMDW-1.0, with users able to select either license

GIFT-Eval results

Foundation models are most useful when they generalize to time series that were not included in task-specific training. GIFT-Eval evaluates forecasting models across heterogeneous datasets and forecasting scenarios.

Among models that are zero-shot, replicable, and evaluated without test leakage, PatchTST-FM-r2 ranked second for both CRPS and MASE as of September 8, 2026. Lower values are better for both metrics.

PatchTST-FM-r2 achieved a geometric-mean CRPS of 0.467, placing it behind TimesFM-3 in the comparison. It achieved a geometric-mean MASE of 0.6846. In the same category, it was the highest-performing model among models with permissive, commercial-friendly licensing.

When pretrained replicable models were also included, PatchTST-FM-r2 ranked third for CRPS and fourth for MASE. It outperformed several pretrained models, including Chronos-2, Timer-S1, and Toto variants, although some competing models were considerably larger.

Architecture changes from PatchTST-FM-r1

PatchTST-FM-r2 retains the patch-based representation used by the PatchTST family, but changes the internal architecture to model long- and short-term relationships more efficiently and smooth predictions between patches.

PatchTST-FM-r1 used standard transformer blocks that combined multi-head self-attention with a feed-forward network. PatchTST-FM-r2 replaces these with conformer-style blocks. Each block contains two half-step feed-forward layers surrounding multi-head self-attention and a temporal convolution layer.

Self-attention models long-range relationships between patches, while convolution supplies an inductive bias toward local temporal structure. This allows the convolution layer to capture shorter-term interactions while attention focuses on relationships across longer horizons. The conformer blocks use alternating convolution kernel sizes of 3 and 5 in the repeating pattern {5, 5, 3, 3}.

The architecture also includes:

50% overlapping patches
Hamming-window weighting
Overlap-and-add forecasting to smooth patch boundaries
Additional normalization for stability
An increase from 20 to 30 blocks

Together, these changes produce a model with approximately 385 million parameters, context lengths of up to 8,192 steps, and predictions across 99 quantiles. The model supports point forecasts as well as quantile outputs for forecasting distributions and uncertainty intervals.

Training data

PatchTST-FM-r2 uses a documented pretraining corpus drawn from four sources:

Selected datasets from GiftEvalPretrain
Custom synthetic data based on KernelSynth, using modified periodic kernels and limited augmentation
A TSMixup corpus generated using the approach described by Chronos, restricted to datasets outside the GIFT-Eval evaluation set
Approximately 500,000 synthetic CauKer sequences, each with a length of 4,096

Documentation about the training corpus helps users assess possible benchmark leakage and consider the model's licensing and governance implications. Organizations still need to conduct their own model-governance and licensing reviews.

Licensing and implementation

Granite Time Series PatchTST-FM-r2 is dual-licensed under Apache 2.0 and OpenMDW 1.0. Users may select either license. Both provide permissive rights to use, modify, and distribute the model, while OpenMDW provides a licensing framework designed for AI models and related materials.

The architecture implementation is available in the Granite-TSFM repository and is backward-compatible with PatchTST-FM-r1 checkpoints.

Python example

The Granite TSFM package can be installed with:

pip install "granite-tsfm>=0.3.9"

The model can then be loaded from the Hugging Face Hub and applied to sample ETTh data:

import pandas as pd
from tsfm_public import PatchTSTFMForPrediction, TimeSeriesForecastingPipeline

## Load model weights
model = PatchTSTFMForPrediction.from_pretrained(
    "ibm-granite/granite-timeseries-patchtst-fm-r2"
)

## Read sample data from ETTh
df = pd.read_csv(
    "https://raw.githubusercontent.com/zhouhaoyi/ETDataset/main/ETT-small/ETTh1.csv",
    parse_dates=["date"],
)

## Set up the forecasting pipeline
pipe = TimeSeriesForecastingPipeline(
    model=model,
    id_columns=[],
    timestamp_column="date",
    target_columns=["HUFL"],
    max_context_length=model.config.context_length,
    context_length=512,
    prediction_length=64,
    impute_method=None,
    quantile_levels=[0.1, 0.5, 0.9],
    explode_forecasts=True,
    freq="1h",
)

## Forecast from the last 512 samples
forecast = pipe(df.iloc[-512:])

The example does not require fine-tuning or task-specific model fitting. The pipeline uses the recent history of the series to generate future forecasts, including the requested quantiles. The input can be replaced with regularly sampled data such as demand, sensor telemetry, CPU utilization, energy consumption, transaction volume, traffic, or prices.

Streaming applications

PatchTST-FM-r2 is part of the broader Granite Time Series model portfolio. IBM and Confluent have also made several Granite Time Series models available through an Early Access program in Confluent Cloud. The initial portfolio includes PatchTST-FM-r1, FlowState-r1.1, TTM-r3, and TSPulse.

The integration provides foundation-model inference for streaming applications through Apache Flink on Confluent Cloud. Forecasts and anomaly-detection results can be generated from live streams instead of requiring data to be moved to a separate machine-learning environment.

IBM Releases Granite Time Series PatchTST-FM-r2 for Zero-Shot Forecasting