IBM 发布了 Granite Time Series PatchTST-FM-r2，它是 Granite TSFM 系列中的最新模型（github、博客）。 PatchTST-FM-r2 是其前身 PatchTST-FM-r1 的新版本，在约 385M 参数模型中结合了更新的架构、更大的预训练语料库、概率预测、对缺失值插补的支持以及强大的零样本性能。
IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license
Hugging Face Blog 发布了这条动态，详情请查看官方发布。
正文 · 来源原文
Time-series foundation models are changing the way forecasting systems are built. Instead of training and maintaining a separate model for every dataset, users can use a pretrained model and generate forecasts zero-shot.
IBM has released Granite Time Series PatchTST-FM-r2, the latest model in the Granite TSFM family (github, blog). PatchTST-FM-r2, a new version of its predecessor PatchTST-FM-r1, combines an updated architecture, a larger pretraining corpus, probabilistic forecasting, support for imputation of missing values, and strong zero-shot performance in a ~385M-parameter model.
As of September 8, 2026, the model is the top performing zero-shot model released under a permissive, commercial-friendly open-source license (Apache 2.0 and OpenMDW 1.0) among replicable, zero-shot models on the GIFT-Eval leaderboard. GIFT-Eval is a comprehensive time series forecasting benchmark designed to evaluate models across diverse forecasting scenarios; the model ranks #2 overall among replicable, zero-shot models.
The model weights, architecture, inference pipeline, and code needed to reproduce the benchmark results are all available.
In this blog we describe the model, dive deeper into the benchmarking results and the model architecture, discuss the training data and licensing, and provide code examples illustrating how to use the model. Finally, we also highlight how the models from the Granite Time Series family can be used in streaming applications in production setting leveraging Confluent product.
Ready to try it? Open Granite Time Series PatchTST-FM-r2 on Hugging Face
General-purpose zero-shot forecasting for demand, prices, energy loads, traffic, telemetry, and other time series. Approximately 385M parameters, context length up to 8,192, flexible forecast lengths, and probabilistic forecasts through a 99-quantile prediction head. The model backbone is constructed from conformer blocks that combine multi-head self-attention with temporal convolution to capture long- and short-range temporal structure. Permissively licensed model with top performance in the replicable zero-shot category of the GIFT-Eval benchmark (dual-licensed under Apache-2.0 and OpenMDW-1.0, users may select either license). Available open weights, architecture, inference pipeline, and code to reproduce the benchmark.
A foundation model is most useful when it generalizes to time series it has not been specifically trained on. For that reason, we focus first on zero-shot performance.
GIFT-Eval provides a broad evaluation of forecasting models across heterogeneous datasets and forecasting scenarios. When restricting the leaderboard to models that are zero-shot, replicable, and evaluated without test leakage, PatchTST-FM-r2 ranks second for both CRPS and MASE as of September 8, 2026, as illustrated in Figures 1 and 2 (lower values are better for both metrics). Importantly, PatchTST-FM-r2 is the highest-performing model in the same category among models with permissive, commercial-friendly licensing.
Figure 1. GIFT-Eval CRPS for leading replicable zero-shot models. PatchTST-FM-r2 achieves a geometric-mean CRPS of 0.467, placing it immediately behind TimesFM-3 in this comparison, and first among the models with permissive licenses.
Figure 2. GIFT-Eval MASE for leading replicable zero-shot models. PatchTST-FM-r2 achieves a geometric-mean MASE of 0.6846. Blue bars denote models released by the IBM time-series foundation-model team.
Some models on GIFT-Eval are categorized as pretrained rather than strictly zero-shot. These models are allowed to include the training portions of GIFT-Eval evaluation datasets in their pretraining corpora.
Even when these pretrained models are added to the comparison, PatchTST-FM-r2 remains near the top as seen in Figures 3 and 4: 3rd for CRPS and 4th for MASE among the replicable models. It outperforms several pretrained models, including Chronos-2, Timer-S1, and Toto variants, despite some competing models being considerably larger.
Figure 3. GIFT-Eval CRPS when both zero-shot and pretrained replicable models are considered.
Figure 4. GIFT-Eval MASE when both zero-shot and pretrained replicable models are considered.
PatchTST-FM-r2 retains the patch-based representation that made the PatchTST family effective, but the internal architecture is redesigned to capture long- and short-term relationships efficiently and to smoothen inter-patch predictions — both of which substantially improve error measures.
One change is the move from standard transformer layers to layers which incorporate convolution along with the multi-head self-attention. These layers are referred to as conformer layers and has its origin in speech processing applications.
Figure 5. Architectural evolution from PatchTST-FM-r1 to the Conformer-based PatchTST-FM-r2.
A PatchTST-FM-r1 block combines multi-head self-attention with a feed-forward network. In r2, we replaced this with a conformer-style block containing two half-step feed-forward layers surrounding multi-head self-attention and a temporal convolution layer.
This gives the model two complementary mechanisms for reasoning over a time series. Self-attention can model long-range relationships between patches, while convolution provides an inductive bias toward local temporal structure. The convolution component can therefore capture shorter-term interactions while allowing attention to concentrate on relationships over longer horizons. This phenomenon is observable in attention patterns captured in the transformer and conformer versions (see example figure below using real samples from the ETTh1 dataset). While a significant portion of the transformer’s self-attention (left in the figure) concentrates near the diagonal, i.e., capturing local relationships, attentions in the conformer block (right in the figure) show long-distance (far-off-diagonal) focus, thanks to the convolution layer covering the short distances. The conformer blocks in the backbone use alternating convolution kernel sizes of 3 and 5, in a repeating pattern {5, 5, 3, 3}.
Figure 6. Comparing attention patterns captured in the transformer (left) and conformer versions(right) using real samples from the ETTh1 dataset.
Additionally, PatchTST-FM-r2 uses 50% overlapping patches with Hamming-window weighting and overlap-and-add forecasting to smooth patch boundaries and improve forecasting accuracy. Finally, the architecture adds normalization for stability and expands from 20 to 30 blocks.
With these changes, the resulting model has approximately 385M parameters, supports very long contexts of up to 8,192 steps, and predicts 99 quantiles over flexible forecast lengths. The model provides both point forecasts and quantile outputs for forecasting distributions and uncertainty intervals.
For foundation models intended for real applications, model quality is only one consideration. Developers increasingly need to understand what data went into a model, whether benchmark data may have leaked into training, and what the implications are for deploying the model.
PatchTST-FM-r2 uses a documented pretraining corpus consisting of 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 by Chronos but restricted to datasets outside the GIFT-Eval evaluation set; and approximately 500,000 synthetic CauKer sequences, each of length 4,096.
For enterprise adopters, this kind of transparency can be as important as another few points on a leaderboard. This does not eliminate the need for an organization's own model-governance and licensing review, but it gives users considerably more information with which to perform that review than an opaque pretraining corpus would.
To provide greater choice for the community, Granite Time Series PatchTST-FM-r2 is dual licensed under Apache 2.0 and OpenMDW 1.0. Users may select either license, both of which provide broad, permissive rights to use, modify, and distribute the models, with the Linux Foundation's OpenMDW offering a licensing framework specifically designed for AI models and related materials. By making the models available under permissive open-source licenses, IBM aims to reduce barriers to adoption and enable organizations, researchers, and developers to build on the technology with confidence that comes with licenses that don't restrict your use.The architecture implementation is also available through Granite-TSFM repository and is backward-compatible with PatchTST-FM-r1 checkpoints.
The easiest way to evaluate a foundation model is on your own time series data.
Install the Granite TSFM package:
Then load PatchTST-FM-r2 directly from the Hugging Face Hub:
As you can see, no fine-tuning and no task-specific model fitting is needed. The pipeline consumes just the recent history of the series and generates the future forecast, including the requested quantiles.
The above is a simple example on publicly available data – you can replace the example input data with your own, including demand, sensor telemetry, CPU utilization, energy consumption, transaction volume, traffic, prices, or another regularly sampled time series.
Try it on your own data: Open PatchTST-FM-r2 on the Hugging Face Hub
This release connects to a broader effort around IBM Granite Time Series models, adding a new model to the broader portfolio.
For applications where data arrives continuously rather than in static DataFrames, IBM and Confluent recently 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 brings foundation-model inference directly into streaming applications through Apache Flink on Confluent Cloud. Forecasts and anomaly-detection results can be generated from live streams rather than requiring teams to set up and move data to a separate ML environment.
