IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license

IBM 发布具备商业友好许可的 SOTA Granite 时间序列模型 PatchTST-FM-r2

High-performance zero-shot forecasting with commercial-friendly open licensing. 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. 具备商业友好开源许可的高性能零样本预测。时间序列基础模型正在改变预测系统的构建方式。用户无需为每个数据集训练和维护单独的模型，而是可以使用预训练模型进行零样本预测。

IBM has released Granite Time Series PatchTST-FM-r2, the latest model in the Granite TSFM family. 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. IBM 发布了 Granite TSFM 系列的最新模型：Granite Time Series PatchTST-FM-r2。作为前代产品 PatchTST-FM-r1 的升级版，PatchTST-FM-r2 结合了更新的架构、更大的预训练语料库、概率预测功能、缺失值填补支持，并在约 3.85 亿参数的模型规模下实现了强大的零样本预测性能。

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. 截至 2026 年 9 月 8 日，在 GIFT-Eval 排行榜的可复现零样本模型中，该模型是采用宽松且商业友好的开源许可（Apache 2.0 和 OpenMDW 1.0）发布的最强模型。GIFT-Eval 是一个全面的时间序列预测基准，旨在评估模型在各种预测场景下的表现；该模型在所有可复现的零样本模型中总排名第二。

TL;DR

简要总结

• General-purpose zero-shot forecasting for demand, prices, energy loads, traffic, telemetry, and other time series.
• 通用零样本预测：适用于需求、价格、能源负载、交通、遥测及其他时间序列数据。
• Approximately 385M parameters, context length up to 8,192, flexible forecast lengths, and probabilistic forecasts through a 99-quantile prediction head.
• 约 3.85 亿参数：上下文长度高达 8,192，支持灵活的预测长度，并通过 99 分位数预测头提供概率预测。
• 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.
• 模型主干：由 Conformer 模块构建，结合了多头自注意力机制与时间卷积，以捕捉长短期时间结构。
• 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).
• 宽松许可：在 GIFT-Eval 基准的可复现零样本类别中表现领先（采用 Apache-2.0 和 OpenMDW-1.0 双重许可）。

Strong zero-shot forecasting on GIFT-Eval

在 GIFT-Eval 上的强大零样本预测表现

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. 基础模型最有价值的地方在于其对未专门训练过的时间序列的泛化能力。因此，我们首先关注零样本性能。GIFT-Eval 在异构数据集和预测场景中对预测模型进行了广泛评估。在仅限零样本、可复现且无测试集泄露的评估中，截至 2026 年 9 月 8 日，PatchTST-FM-r2 在 CRPS 和 MASE 指标上均排名第二。

Architecture: What changed from PatchTST-FM-r1?

架构：相较于 PatchTST-FM-r1 有何变化？

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. 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. PatchTST-FM-r2 保留了使 PatchTST 系列高效的基于补丁（patch-based）的表示方法，但重新设计了内部架构，以更高效地捕捉长短期关系并平滑补丁间的预测。其中一个变化是从标准的 Transformer 层转向了结合卷积与多头自注意力的层。这些层被称为 Conformer 层，起源于语音处理应用。

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. PatchTST-FM-r1 模块结合了多头自注意力和前馈网络。在 r2 版本中，我们将其替换为 Conformer 风格的模块，包含两个半步前馈层，环绕着多头自注意力和时间卷积层。这为模型提供了两种互补的时间序列推理机制：自注意力机制可以建模补丁之间的长距离关系，而卷积则为局部时间结构提供了归纳偏置。