IBM’s PatchTST-FM-r2 Tops Open-Source Zero-Shot Forecasting Leaderboard

IBM Granite Time Series PatchTST-FM-r2 Breaks the Zero-Shot License Barrier

IBM has introduced Granite Time Series PatchTST-FM-r2, a roughly 385-million-parameter foundation model built for zero-shot forecasting across demand, energy loads, traffic, telemetry, and prices.

On September 8, 2026, the model became the top-performing zero-shot system released under a permissive, commercially usable open-source license on the GIFT-Eval leaderboard.

It ranks second overall among replicable zero-shot models, immediately behind Google Research’s TimesFM-3, while holding the lead among models that do not restrict commercial use.

The release pairs a redesigned conformer-based architecture with a 99-quantile prediction head, missing-value imputation, and a context window of up to 8,192 time steps, as detailed in the Hugging Face blog.

Dual licensing under Apache 2.0 and OpenMDW 1.0 lets users choose either framework and removes a major adoption barrier for production deployments.

Inside the Conformer Architecture and Quantile Forecasting Upgrade

The most substantial change from PatchTST-FM-r1 is the move from standard transformer layers to conformer-style blocks.

A conformer block surrounds multi-head self-attention with two half-step feed-forward layers and inserts a temporal convolution module.

Self-attention captures long-range relationships between patches, while the convolution path adds an inductive bias toward local temporal structure.

That division of labor allows attention to focus on longer horizons because convolution handles shorter-range interactions.

The backbone uses alternating convolution kernel sizes of 3 and 5 in a repeating pattern of 5, 5, 3, 3.

IBM also moved from 20 to 30 blocks, added normalization for stability, and introduced 50% overlapping patches with Hamming-window weighting.

Overlap-and-add forecasting smooths patch boundaries and improves error measures.

Key technical specifications include:

• Context length: Up to 8,192 steps.
• Forecast outputs: Point forecasts and 99 quantiles for uncertainty intervals.
• Patch design: 50% overlap with Hamming-window weighting.
• Backbone depth: Expanded to 30 blocks.

The pretraining corpus is documented rather than opaque.

It draws from selected GiftEvalPretrain datasets, KernelSynth-based synthetic data with modified periodic kernels, a TSMixup corpus restricted to datasets outside the GIFT-Eval evaluation set, and roughly 500,000 synthetic CauKer sequences of length 4,096.

That transparency does not replace a company’s own model-governance review, but it gives enterprise adopters far more information for that process.

OpenMDW 1.0, created by the Linux Foundation, provides a licensing framework specifically designed for AI models and related materials.

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

Teams can install the Granite-TSFM package version 0.3.9 or newer and load the model directly from Hugging Face Hub.

The pipeline consumes recent series history and generates forecasts with requested quantiles, bypassing fine-tuning and task-specific fitting entirely.

Beyond static notebooks, IBM and Confluent have made several Granite Time Series models available through an Early Access program in Confluent Cloud.

The integration brings foundation-model inference into streaming applications through Apache Flink on Confluent Cloud.

The Early Access portfolio currently includes PatchTST-FM-r1, FlowState-r1.1, TTM-r3, and TSPulse, with PatchTST-FM-r2 arriving as the newest addition.

Forecasts and anomaly-detection results can be generated from live streams instead of requiring separate ML environments.

Where IBM Lands in a Crowded Time-Series Foundation-Model Market

PatchTST-FM-r2 enters a field where Google Research’s TimesFM-3 sits at the top of the GIFT-Eval leaderboard for pretrained foundation models.

According to Google Research, TimesFM-3 is a 330-million-parameter decoder-only transformer pretrained on more than one trillion real and synthetic time points.

That model is natively multivariate, supports multiple targets and covariates, and predicts nine quantiles from the 10th to 90th percentile for each target series.

IBM’s differentiator lies in dense 99-quantile output and a permissive dual license, not in matching TimesFM-3’s multivariate coverage point for point.

Even when the comparison includes pretrained models allowed to use GIFT-Eval training portions, PatchTST-FM-r2 places third for CRPS and fourth for MASE among replicable models.

It outperforms several pretrained systems, including Chronos-2, Timer-S1, and Toto variants, despite some competitors being considerably larger.

Independent benchmark evidence from a peer-reviewed FETS study in Energy and AI adds a useful caveat.

Across 54 energy datasets, Chronos-2 achieved the lowest overall median NRMSE of 0.472, with TiRex-2 close behind at 0.474.

Those foundation models outperformed task-specific XGBoost and random forest baselines at 0.611 and 0.696, respectively.

PatchTST-FM-r2 was not included in that energy-specific study, so its domain performance outside GIFT-Eval remains unverified at the independent benchmark level.

Because the weights, architecture, inference pipeline, and benchmark reproduction code are open, the GIFT-Eval numbers can be independently reproduced.

For enterprise buyers, the competitive tension is clear: IBM has not claimed the outright top spot on GIFT-Eval, but it owns the strongest permissive-license position today.

That distinction may matter more in procurement than a narrow gap in CRPS.

The Forecast for Enterprise Forecasting Itself

For enterprise teams, this release resets the baseline: high-fidelity zero-shot forecasting no longer requires an opaque pretraining corpus or a restrictive license. For teams building AI-powered forecasting pipelines that need to scale, programmatic SEO AI automation is how Andres SEO Expert approaches it — contact Andres SEO Expert.