IBM's Granite PatchTST-FM-r2 Tops Permissively Licensed Time Series Forecasting Models on GIFT-Eval

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

IBM has released Granite Time Series PatchTST-FM-r2, a ~385 million-parameter foundation model for time series forecasting, according to an IBM Research blog post published September 9, 2026. As of September 8, 2026, the model ranks second overall among replicable, zero-shot models on the GIFT-Eval benchmark — and is the top-performing model in that category among those released under a permissive, commercial-friendly open-source license, according to IBM.

The model is dual-licensed under Apache-2.0 and OpenMDW-1.0, letting users choose either license. Weights, architecture, inference pipeline, and the code needed to reproduce the benchmark results are all publicly available on Hugging Face.

Benchmark Results

On GIFT-Eval, a benchmark designed to test forecasting models across heterogeneous datasets and scenarios, PatchTST-FM-r2 achieves a geometric-mean CRPS (continuous ranked probability score, lower is better) of 0.467 and a geometric-mean MASE (mean absolute scaled error, lower is better) of 0.6846. IBM states this places the model immediately behind TimesFM-3 among replicable zero-shot models on both metrics, and first among models with permissive licenses.

When IBM expanded the comparison to include "pretrained" models — those allowed to include GIFT-Eval training-set data during pretraining — PatchTST-FM-r2 still ranks third for CRPS and fourth for MASE among replicable models, according to the company. IBM claims it outperforms Chronos-2, Timer-S1, and Toto variants in this expanded comparison, despite some competing models being larger.

Architecture Changes from r1

PatchTST-FM-r2 replaces the standard transformer layers used in predecessor PatchTST-FM-r1 with "conformer" blocks — a design borrowed from speech processing that combines multi-head self-attention with temporal convolution. Each block contains two half-step feed-forward layers surrounding a self-attention layer and a temporal convolution layer, using alternating kernel sizes of 3 and 5 in a repeating {5, 5, 3, 3} pattern.

According to IBM, this combination lets attention layers focus on long-range dependencies between patches while convolution handles local, short-range temporal structure — a pattern the company says is visible in attention maps from tests on the ETTh1 dataset. The architecture also adds 50% overlapping patches with Hamming-window weighting and overlap-and-add forecasting to smooth predictions across patch boundaries, additional normalization layers for training stability, and an increase from 20 to 30 blocks total.

Specifications

The model supports context lengths up to 8,192 time steps and flexible forecast horizons. It produces both point forecasts and probabilistic outputs via a 99-quantile prediction head, and supports imputation of missing values in input sequences. IBM says the model is intended for general-purpose forecasting across domains including demand planning, pricing, energy load, traffic, and telemetry data.

Training data draws from four documented sources: selected datasets from GiftEvalPretrain, custom synthetic data generated using a modified KernelSynth approach, a TSMixup-style corpus following the method described by Chronos, and additional unspecified real-world data. IBM did not disclose exact dataset sizes or a training cutoff date in the published material.

IBM also highlighted integration with Confluent for using Granite Time Series models in streaming production environments, though further technical details were not provided.

What This Means

PatchTST-FM-r2 is notable less for topping the overall leaderboard — it doesn't — and more for closing the gap between permissively licensed open models and the best-performing systems on GIFT-Eval, some of which carry restrictive licenses or undisclosed training data. For enterprises building forecasting pipelines for inventory, energy, or infrastructure telemetry, a 385M-parameter model with Apache-2.0/OpenMDW licensing and reproducible benchmark code removes a common blocker: uncertainty over whether a state-of-the-art forecasting model can be deployed commercially without licensing risk. The architectural shift to conformer blocks — borrowing directly from speech recognition — also signals that time series foundation models are converging on hybrid attention-convolution designs already proven in other sequence-modeling domains, rather than defaulting to pure transformer architectures.