{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-09-28T06:03:00.468Z","headline":"IBM 发布 Granite Time Series PatchTST-FM-r2，采用 Apache 2.0 与 OpenMDW 1.0 双许可","description":"IBM 发布 Granite Time Series PatchTST-FM-r2，约 385M 参数，支持最长 8，192 上下文、99 分位数概率预测和缺失值插补。","url":"https://www.aioga.com/news/cmtuaa1fw14xzrofpxqrrz1t3/","mainEntityOfPage":"https://www.aioga.com/news/cmtuaa1fw14xzrofpxqrrz1t3/","datePublished":"2026-09-09T15:36:24.000Z","dateModified":"2026-09-09T15:36:24.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://huggingface.co/blog/ibm-research/ibm-releases-sota-granite-time-series","https://aihot.news/items/cmtuaa1fw14xzrofpxqrrz1t3"],"canonicalUrl":"https://www.aioga.com/news/cmtuaa1fw14xzrofpxqrrz1t3/","directAnswer":{"@type":"Answer","text":"Aioga 编辑摘要：IBM 发布 Granite Time Series PatchTST-FM-r2，约 385M 参数，支持最长 8，192 上下文、99 分位数概率预测和缺失值插补。 Aioga 将其归入「行业动态」方向，重点关注它对真实使用和行业竞争的影响。","url":"https://www.aioga.com/news/cmtuaa1fw14xzrofpxqrrz1t3/","dateCreated":"2026-09-09T15:36:24.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"huggingface.co source article","url":"https://huggingface.co/blog/ibm-research/ibm-releases-sota-granite-time-series","datePublished":"2026-09-09T15:36:24.000Z","provider":{"@type":"Organization","name":"huggingface.co","url":"https://huggingface.co/blog/ibm-research/ibm-releases-sota-granite-time-series"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.news/items/cmtuaa1fw14xzrofpxqrrz1t3","datePublished":"2026-09-09T15:36:24.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.news/items/cmtuaa1fw14xzrofpxqrrz1t3"}}],"aggregationSource":"Hugging Face：Blog（RSS）","originalPublisher":{"name":"huggingface.co","url":"https://huggingface.co/blog/ibm-research/ibm-releases-sota-granite-time-series"},"geoDeepAnswer":null,"article":{"id":"cmtuaa1fw14xzrofpxqrrz1t3","slug":"cmtuaa1fw14xzrofpxqrrz1t3","url":"https://www.aioga.com/news/cmtuaa1fw14xzrofpxqrrz1t3/","title":"IBM 发布 Granite Time Series PatchTST-FM-r2，采用 Apache 2.0 与 OpenMDW 1.0 双许可","title_en":"","summary":"IBM 发布 Granite Time Series PatchTST-FM-r2，约 385M 参数，支持最长 8，192 上下文、99 分位数概率预测和缺失值插补。","source":"Hugging Face：Blog（RSS）","sourceUrl":"https://huggingface.co/blog/ibm-research/ibm-releases-sota-granite-time-series","aiHotUrl":"https://aihot.news/items/cmtuaa1fw14xzrofpxqrrz1t3","publishedAt":"2026-09-09T15:36:24.000Z","category":"行业动态","score":58,"selected":false,"articleBody":["TL;DR ：#tldr Strong zero-shot forecasting on GIFT-Eval ：#strong-zero-shot-forecasting-on-gift-eval Competitive even against models allowed to use benchmark training data ：#competitive-even-against-models-allowed-to-use-benchmark-training-data Architecture: What changed from PatchTST-FM-r1? ：#architecture-what-changed-from-patchtst-fm-r1 Training data ：#training-data Open for research experimentation — and for commercial use ：#open-for-research-experimentation--and-for-commercial-use Try PatchTST-FM-r2 in a few lines of Python ：#try-patchtst-fm-r2-in-a-few-lines-of-python From notebooks to streaming time series ：#from-notebooks-to-streaming-time-series 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.","IBM has released Granite Time Series PatchTST-FM-r2 ：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r2, the latest model in the Granite TSFM family (github：https://github.com/ibm-granite/granite-tsfm, blog：https://research.ibm.com/blog/time-series-ai-enterprise). PatchTST-FM-r2, a new version of its predecessor PatchTST-FM-r1：https://huggingface.co/ibm-granite/granite-timeseries-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：https://events.confluent.io/early-access-flink-features.","Ready to try it? Open Granite Time Series PatchTST-FM-r2 on Hugging Face ：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r2","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：https://arxiv.org/pdf/2005.08100 in speech processing applications.","：https://cdn-uploads.huggingface.co/production/uploads/64b01b5252349201f972ba17/gYPoFfbdOqaT9nZmABON1.png","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：https://www.apache.org/licenses/LICENSE-2.0.txt and OpenMDW 1.0：https://raw.githubusercontent.com/OpenMDW/OpenMDW/refs/heads/main/1.0/LICENSE.openmdw. 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：https://github.com/ibm-granite/granite-tsfm and is backward-compatible with PatchTST-FM-r1 checkpoints.","The easiest way to evaluate a foundation model is on your own time series data.","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 ：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r2","This release connects to a broader effort around IBM Granite Time Series models, adding a new model to the broader portfolio：https://research.ibm.com/blog/time-series-ai-enterprise.","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."],"articleImages":[{"sourceUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1627505688463-60107b385ac3e86b3ea4fc34.jpeg","alt":"","afterParagraph":0,"url":"/media/articles/cmtuaa1fw14xzrofpxqrrz1t3/5f7b8d2bdd3ff09f.webp"},{"sourceUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64e8143f6de557454220921e/aNouKBrKm1dnN9pI1JXlp.jpeg","alt":"","afterParagraph":0,"url":"/media/articles/cmtuaa1fw14xzrofpxqrrz1t3/f257760baa1a5507.webp"},{"sourceUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/vy5ixwWBqZAnGxVVtKx9m.png","alt":"","afterParagraph":0,"url":"/media/articles/cmtuaa1fw14xzrofpxqrrz1t3/4b3b401904877f65.webp"}],"mediaStatus":"ok","articleBodyZh":["简要总结：#tldr 在 GIFT-Eval 上实现强大的零样本预测 #strong-zero-shot-forecasting-on-gift-eval 即使与获准使用基准训练数据的模型相比，也具有竞争力 #competitive-even-against-models-allowed-to-use-benchmark-training-data 架构：与 PatchTST-FM-r1 相比有哪些变化？ #architecture-what-changed-from-patchtst-fm-r1 训练数据 #training-data 开放用于研究实验，也可用于商业用途 #open-for-research-experimentation--and-for-commercial-use 只需几行 Python 代码即可试用 PatchTST-FM-r2 #try-patchtst-fm-r2-in-a-few-lines-of-python 从笔记本到流式时间序列 #from-notebooks-to-streaming-time-series 采用对商业友好的开放许可协议，实现高性能零样本预测。","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并进行零样本预测。","IBM 发布了 Granite Time Series PatchTST-FM-r2：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r2，这是 Granite TSFM 系列的最新模型（GitHub：https://github.com/ibm-granite/granite-tsfm，博客：https://research.ibm.com/blog/time-series-ai-enterprise）。PatchTST-FM-r2 是其前身 PatchTST-FM-r1 的新版本：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r1，结合了更新的架构、更大的预训练语料库、概率预测、缺失值插补支持，以及在约 3.85 亿参数模型中实现的强大零样本性能。","截至 2026 年 9 月 8 日，该模型是 GIFT-Eval 排行榜上可复现的零样本模型中性能最顶尖、且基于宽松商业友好开源许可（Apache 2.0 和 OpenMDW 1.0）发布的模型。GIFT-Eval 是一个综合的时间序列预测基准，用于评估模型在各种预测场景中的表现；该模型在可复现零样本模型中整体排名第二。","模型权重、架构、推理管道以及复现基准结果所需的代码均可获得。","在本博客中，我们将描述模型，深入探讨基准测试结果和模型架构，讨论训练数据和许可问题，并提供代码示例说明如何使用该模型。最后，我们还将强调 Granite 时间序列系列模型如何在生产环境中利用 Confluent 产品用于流处理应用：https://events.confluent.io/early-access-flink-features。","准备好尝试了吗？在 Hugging Face 上打开 Granite 时间序列 PatchTST-FM-r2：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r2","一个基础模型最有用的情况是其能够泛化到未被专门训练的时间序列。出于这个原因，我们首先关注零样本性能。","GIFT-Eval 提供了对异构数据集和预测场景下的预测模型的广泛评估。在将排行榜限制为零样本、可复现且在无测试泄漏情况下评估的模型时，截至 2026 年 9 月 8 日，PatchTST-FM-r2 在 CRPS 和 MASE 上均排名第二，如图 1 和图 2 所示（两个指标的值越低越好）。重要的是，PatchTST-FM-r2 在具有宽松、商业友好许可的同类别模型中表现最佳。","图 1. GIFT-Eval 可复现零样本模型的 CRPS。PatchTST-FM-r2 的几何平均 CRPS 为 0.467，在此比较中紧随 TimesFM-3 之后，并在具有宽松许可的模型中排名第一。","图 2. GIFT-Eval 可复现零样本模型的 MASE。PatchTST-FM-r2 的几何平均 MASE 为 0.6846。蓝色柱表示由 IBM 时间序列基础模型团队发布的模型。","GIFT-Eval 上的一些模型被归类为预训练模型，而非严格的零样本模型。这些模型被允许在其预训练语料中包含 GIFT-Eval 评估数据集的训练部分。","即使这些预训练模型被加入比较，PatchTST-FM-r2 仍然位居前列，如图 3 和图 4 所示：在可复现模型中 CRPS 排名第 3，MASE 排名第 4。它优于多个预训练模型，包括 Chronos-2、Timer-S1 及 Toto 变体，尽管一些竞争模型规模更大。","图3. 在同时考虑零样本和预训练可复现模型时的 GIFT-Eval CRPS。","图4. 在同时考虑零样本和预训练可复现模型时的 GIFT-Eval MASE。","PatchTST-FM-r2 保留了使 PatchTST 系列有效的基于 patch 的表示，但内部架构经过重新设计，以高效捕捉长短期关系并平滑补丁间的预测——这两者都显著改善了误差指标。","一个变化是从标准 transformer 层转向结合卷积和多头自注意力的层。这些层被称为 conformer 层，其起源：https://arxiv.org/pdf/2005.08100，最初用于语音处理应用。","：https://cdn-uploads.huggingface.co/production/uploads/64b01b5252349201f972ba17/gYPoFfbdOqaT9nZmABON1.png","图 5. 从 PatchTST-FM-r1 到基于 Conformer 的 PatchTST-FM-r2 的架构演进。","一个 PatchTST-FM-r1 模块将多头自注意力与前馈网络结合。在 r2 中，我们用一个 conformer 风格的模块替换了它，该模块包含两个半步前馈层，包围多头自注意力层和一个时序卷积层。","这为模型提供了两种互补的时间序列推理机制。自注意力可以建模 patch 之间的长程关系，而卷积为局部时间结构提供归纳偏置。因此，卷积部分可以捕捉短期交互，同时允许注意力集中于较长时间跨度的关系。这一现象可以在 transformer 和 conformer 版本的注意力模式中观察到（见下图使用 ETTh1 数据集的真实样本举例）。虽然 transformer 的大部分自注意力（图中左侧）集中在对角线附近，即捕捉局部关系，但在 conformer 模块中的注意力（图中右侧）显示出远距离（对角线之外）的关注，这归功于卷积层覆盖了短距离。骨干网络中的 conformer 模块使用 3 和 5 的交替卷积核尺寸，以重复模式 {5, 5, 3, 3} 进行交替。","图 6. 使用来自 ETTh1 数据集的真实样本比较变压器（左）和卷积变体（右）捕获的注意力模式。","此外，PatchTST-FM-r2 使用 50% 重叠的补丁，配合汉明窗加权和重叠相加预测方法，以平滑补丁边界并提高预测准确性。最后，该架构增加了归一化以增强稳定性，并将块数从 20 增加到 30 块。","通过这些更改，得到的模型大约有 3.85 亿参数，支持最长 8,192 步的长上下文，并预测灵活预测长度下的 99 个分位数。该模型同时提供点预测和分位数输出，用于预测分布和不确定区间。","对于面向实际应用的基础模型，模型质量只是考虑的一个方面。开发者越来越需要了解模型使用了哪些数据，基准数据是否可能泄露到训练中，以及部署该模型的潜在影响。","PatchTST-FM-r2 使用了记录明确的预训练语料库，包含四个来源：来自 GiftEvalPretrain 的精选数据集；基于 KernelSynth 修改周期核和有限增强生成的自定义合成数据；使用 Chronos 描述的方法生成的 TSMixup 语料，但仅限于 GIFT-Eval 评估集之外的数据集；以及大约 50 万个长度为 4,096 的合成 CauKer 序列。","对于企业用户来说，这种透明度可能和在排行榜上多获得几分同样重要。这并不消除组织自身进行模型治理和许可审查的必要性，但与不透明的预训练语料库相比，它为用户提供了更多信息以进行该审查。","为了为社区提供更多选择，Granite 时间序列 PatchTST-FM-r2 采用 Apache 2.0 许可证（https://www.apache.org/licenses/LICENSE-2.0.txt）和 OpenMDW 1.0 许可证（https://raw.githubusercontent.com/OpenMDW/OpenMDW/refs/heads/main/1.0/LICENSE.openmdw）双重授权。用户可以选择任一许可证，这两种许可证都提供广泛、宽松的使用、修改和分发模型的权利，其中 Linux 基金会的 OpenMDW 提供了专为 AI 模型及相关材料设计的许可框架。通过在宽松的开源许可证下提供模型，IBM 旨在降低采纳门槛，使组织、研究人员和开发者能够在不受使用限制的许可证下放心地基于技术构建。该架构实现同样通过 Granite-TSFM 仓库提供（https://github.com/ibm-granite/granite-tsfm），并向后兼容 PatchTST-FM-r1 检查点。","评估基础模型最简单的方法是使用您自己的时间序列数据。","然后直接从 Hugging Face Hub 加载 PatchTST-FM-r2：","如您所见，无需微调，也无需针对特定任务进行模型拟合。该流水线只使用时间序列的近期历史并生成未来预测，包括所需的分位数。","上述示例基于公开数据——您可以用自己的数据替换示例输入数据，包括需求、传感器遥测、CPU 利用率、能耗、交易量、交通量、价格或其他定期采样的时间序列。","在您自己的数据上试试：在 Hugging Face Hub 上打开 PatchTST-FM-r2：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r2","此次发布连接到 IBM Granite 时间序列模型的更广泛努力，为更广泛的组合增加了新模型：https://research.ibm.com/blog/time-series-ai-enterprise","对于数据连续到达而非静态 DataFrame 的应用，IBM 和 Confluent 最近通过 Confluent Cloud 的提前访问计划提供了多个 Granite 时间序列模型。初始组合包括 PatchTST-FM-r1、FlowState-r1.1、TTM-r3 和 TSPulse。","该集成通过 Confluent Cloud 上的 Apache Flink 将基础模型推理直接引入流式应用程序。预测和异常检测结果可以从实时流中生成，而无需团队设置并将数据移动到单独的机器学习环境中。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Aioga 编辑摘要：IBM 发布 Granite Time Series PatchTST-FM-r2，约 385M 参数，支持最长 8，192 上下文、99 分位数概率预测和缺失值插补。 Aioga 将其归入「行业动态」方向，重点关注它对真实使用和行业竞争的影响。","background":"背景分析：公司与行业类动态需要放在竞争格局、商业化路径、资本信号和监管环境中观察，单条公告不能代表最终结果。","viewpoint":"Aioga 判断：这条动态更适合作为行业观察信号，当前信息足以建立线索，但不足以推导长期结论。","implications":"影响分析：对相关团队而言，短期应先核对来源、可用范围和实际成本，再判断是否值得接入或跟进。","nextStep":"后续观察：继续观察官方文件、合作落地、收入或用户信号、竞品动作和监管后续。","evidenceRefs":["title","summary","articleBody"],"confidence":"medium","status":"published","aiGenerated":false,"autoApproved":true,"generatedBy":"rule-safe-fallback","generatedAt":"2026-09-28T06:19:15.330Z","sourceHash":"8033f8c47b46219b","validation":{"passed":true,"mode":"rule-safe-fallback","checks":["schema","length","source-attribution","no-html"]}},"tags":["行业动态","Hugging Face：Blog（RSS）"],"translations":{"zh-CN":{"title":"IBM 发布 Granite Time Series PatchTST-FM-r2，采用 Apache 2.0 与 OpenMDW 1.0 双许可","summary":"IBM 发布 Granite Time Series PatchTST-FM-r2，约 385M 参数，支持最长 8，192 上下文、99 分位数概率预测和缺失值插补。","category":"行业动态","source":"huggingface.co","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM 发布 Granite Time Series PatchTST-FM-r2，采用 Apache 2.0 与 OpenMDW 1.0 双许可 - Aioga AI资讯","description":"IBM 发布 Granite Time Series PatchTST-FM-r2，约 385M 参数，支持最长 8，192 上下文、99 分位数概率预测和缺失值插补。","url":"https://www.aioga.com/news/cmtuaa1fw14xzrofpxqrrz1t3/","articleBody":["简要总结：#tldr 在 GIFT-Eval 上实现强大的零样本预测 #strong-zero-shot-forecasting-on-gift-eval 即使与获准使用基准训练数据的模型相比，也具有竞争力 #competitive-even-against-models-allowed-to-use-benchmark-training-data 架构：与 PatchTST-FM-r1 相比有哪些变化？ #architecture-what-changed-from-patchtst-fm-r1 训练数据 #training-data 开放用于研究实验，也可用于商业用途 #open-for-research-experimentation--and-for-commercial-use 只需几行 Python 代码即可试用 PatchTST-FM-r2 #try-patchtst-fm-r2-in-a-few-lines-of-python 从笔记本到流式时间序列 #from-notebooks-to-streaming-time-series 采用对商业友好的开放许可协议，实现高性能零样本预测。","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并进行零样本预测。","IBM 发布了 Granite Time Series PatchTST-FM-r2：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r2，这是 Granite TSFM 系列的最新模型（GitHub：https://github.com/ibm-granite/granite-tsfm，博客：https://research.ibm.com/blog/time-series-ai-enterprise）。PatchTST-FM-r2 是其前身 PatchTST-FM-r1 的新版本：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r1，结合了更新的架构、更大的预训练语料库、概率预测、缺失值插补支持，以及在约 3.85 亿参数模型中实现的强大零样本性能。","截至 2026 年 9 月 8 日，该模型是 GIFT-Eval 排行榜上可复现的零样本模型中性能最顶尖、且基于宽松商业友好开源许可（Apache 2.0 和 OpenMDW 1.0）发布的模型。GIFT-Eval 是一个综合的时间序列预测基准，用于评估模型在各种预测场景中的表现；该模型在可复现零样本模型中整体排名第二。","模型权重、架构、推理管道以及复现基准结果所需的代码均可获得。","在本博客中，我们将描述模型，深入探讨基准测试结果和模型架构，讨论训练数据和许可问题，并提供代码示例说明如何使用该模型。最后，我们还将强调 Granite 时间序列系列模型如何在生产环境中利用 Confluent 产品用于流处理应用：https://events.confluent.io/early-access-flink-features。","准备好尝试了吗？在 Hugging Face 上打开 Granite 时间序列 PatchTST-FM-r2：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r2","一个基础模型最有用的情况是其能够泛化到未被专门训练的时间序列。出于这个原因，我们首先关注零样本性能。","GIFT-Eval 提供了对异构数据集和预测场景下的预测模型的广泛评估。在将排行榜限制为零样本、可复现且在无测试泄漏情况下评估的模型时，截至 2026 年 9 月 8 日，PatchTST-FM-r2 在 CRPS 和 MASE 上均排名第二，如图 1 和图 2 所示（两个指标的值越低越好）。重要的是，PatchTST-FM-r2 在具有宽松、商业友好许可的同类别模型中表现最佳。","图 1. GIFT-Eval 可复现零样本模型的 CRPS。PatchTST-FM-r2 的几何平均 CRPS 为 0.467，在此比较中紧随 TimesFM-3 之后，并在具有宽松许可的模型中排名第一。","图 2. GIFT-Eval 可复现零样本模型的 MASE。PatchTST-FM-r2 的几何平均 MASE 为 0.6846。蓝色柱表示由 IBM 时间序列基础模型团队发布的模型。","GIFT-Eval 上的一些模型被归类为预训练模型，而非严格的零样本模型。这些模型被允许在其预训练语料中包含 GIFT-Eval 评估数据集的训练部分。","即使这些预训练模型被加入比较，PatchTST-FM-r2 仍然位居前列，如图 3 和图 4 所示：在可复现模型中 CRPS 排名第 3，MASE 排名第 4。它优于多个预训练模型，包括 Chronos-2、Timer-S1 及 Toto 变体，尽管一些竞争模型规模更大。","图3. 在同时考虑零样本和预训练可复现模型时的 GIFT-Eval CRPS。","图4. 在同时考虑零样本和预训练可复现模型时的 GIFT-Eval MASE。","PatchTST-FM-r2 保留了使 PatchTST 系列有效的基于 patch 的表示，但内部架构经过重新设计，以高效捕捉长短期关系并平滑补丁间的预测——这两者都显著改善了误差指标。","一个变化是从标准 transformer 层转向结合卷积和多头自注意力的层。这些层被称为 conformer 层，其起源：https://arxiv.org/pdf/2005.08100，最初用于语音处理应用。","：https://cdn-uploads.huggingface.co/production/uploads/64b01b5252349201f972ba17/gYPoFfbdOqaT9nZmABON1.png","图 5. 从 PatchTST-FM-r1 到基于 Conformer 的 PatchTST-FM-r2 的架构演进。","一个 PatchTST-FM-r1 模块将多头自注意力与前馈网络结合。在 r2 中，我们用一个 conformer 风格的模块替换了它，该模块包含两个半步前馈层，包围多头自注意力层和一个时序卷积层。","这为模型提供了两种互补的时间序列推理机制。自注意力可以建模 patch 之间的长程关系，而卷积为局部时间结构提供归纳偏置。因此，卷积部分可以捕捉短期交互，同时允许注意力集中于较长时间跨度的关系。这一现象可以在 transformer 和 conformer 版本的注意力模式中观察到（见下图使用 ETTh1 数据集的真实样本举例）。虽然 transformer 的大部分自注意力（图中左侧）集中在对角线附近，即捕捉局部关系，但在 conformer 模块中的注意力（图中右侧）显示出远距离（对角线之外）的关注，这归功于卷积层覆盖了短距离。骨干网络中的 conformer 模块使用 3 和 5 的交替卷积核尺寸，以重复模式 {5, 5, 3, 3} 进行交替。","图 6. 使用来自 ETTh1 数据集的真实样本比较变压器（左）和卷积变体（右）捕获的注意力模式。","此外，PatchTST-FM-r2 使用 50% 重叠的补丁，配合汉明窗加权和重叠相加预测方法，以平滑补丁边界并提高预测准确性。最后，该架构增加了归一化以增强稳定性，并将块数从 20 增加到 30 块。","通过这些更改，得到的模型大约有 3.85 亿参数，支持最长 8,192 步的长上下文，并预测灵活预测长度下的 99 个分位数。该模型同时提供点预测和分位数输出，用于预测分布和不确定区间。","对于面向实际应用的基础模型，模型质量只是考虑的一个方面。开发者越来越需要了解模型使用了哪些数据，基准数据是否可能泄露到训练中，以及部署该模型的潜在影响。","PatchTST-FM-r2 使用了记录明确的预训练语料库，包含四个来源：来自 GiftEvalPretrain 的精选数据集；基于 KernelSynth 修改周期核和有限增强生成的自定义合成数据；使用 Chronos 描述的方法生成的 TSMixup 语料，但仅限于 GIFT-Eval 评估集之外的数据集；以及大约 50 万个长度为 4,096 的合成 CauKer 序列。","对于企业用户来说，这种透明度可能和在排行榜上多获得几分同样重要。这并不消除组织自身进行模型治理和许可审查的必要性，但与不透明的预训练语料库相比，它为用户提供了更多信息以进行该审查。","为了为社区提供更多选择，Granite 时间序列 PatchTST-FM-r2 采用 Apache 2.0 许可证（https://www.apache.org/licenses/LICENSE-2.0.txt）和 OpenMDW 1.0 许可证（https://raw.githubusercontent.com/OpenMDW/OpenMDW/refs/heads/main/1.0/LICENSE.openmdw）双重授权。用户可以选择任一许可证，这两种许可证都提供广泛、宽松的使用、修改和分发模型的权利，其中 Linux 基金会的 OpenMDW 提供了专为 AI 模型及相关材料设计的许可框架。通过在宽松的开源许可证下提供模型，IBM 旨在降低采纳门槛，使组织、研究人员和开发者能够在不受使用限制的许可证下放心地基于技术构建。该架构实现同样通过 Granite-TSFM 仓库提供（https://github.com/ibm-granite/granite-tsfm），并向后兼容 PatchTST-FM-r1 检查点。","评估基础模型最简单的方法是使用您自己的时间序列数据。","然后直接从 Hugging Face Hub 加载 PatchTST-FM-r2：","如您所见，无需微调，也无需针对特定任务进行模型拟合。该流水线只使用时间序列的近期历史并生成未来预测，包括所需的分位数。","上述示例基于公开数据——您可以用自己的数据替换示例输入数据，包括需求、传感器遥测、CPU 利用率、能耗、交易量、交通量、价格或其他定期采样的时间序列。","在您自己的数据上试试：在 Hugging Face Hub 上打开 PatchTST-FM-r2：https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r2","此次发布连接到 IBM Granite 时间序列模型的更广泛努力，为更广泛的组合增加了新模型：https://research.ibm.com/blog/time-series-ai-enterprise","对于数据连续到达而非静态 DataFrame 的应用，IBM 和 Confluent 最近通过 Confluent Cloud 的提前访问计划提供了多个 Granite 时间序列模型。初始组合包括 PatchTST-FM-r1、FlowState-r1.1、TTM-r3 和 TSPulse。","该集成通过 Confluent Cloud 上的 Apache Flink 将基础模型推理直接引入流式应用程序。预测和异常检测结果可以从实时流中生成，而无需团队设置并将数据移动到单独的机器学习环境中。"]},"en":{"title":"IBM released Granite Time Series PatchTST-FM-r2, which uses both Apache 2.0 and OpenMDW 1.0 licenses","summary":"IBM released Granite Time Series PatchTST-FM-r2, with approximately 385M parameters, supporting up to 8,192 contexts, 99-quantile probability prediction, and missing value interpolation.","category":"Industry","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM released Granite Time Series PatchTST-FM-r2, which uses both Apache 2.0 and OpenMDW 1.0 licenses - Aioga AI News","description":"IBM released Granite Time Series PatchTST-FM-r2, with approximately 385M parameters, supporting up to 8,192 contexts, 99-quantile probability prediction, and missing value interpol...","url":"https://www.aioga.com/en/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:47:54.852Z"},"ja":{"title":"IBMはApache 2.0とOpenMDW 1.0の両方のライセンスを使用するGranite Time Series PatchTST-FM-r2をリリースしました","summary":"IBMは約3億8千5百万パラメータを持ち、最大8,192のコンテキスト、99分位数確率予測、欠損値補間に対応したGranite Time Series PatchTST-FM-r2をリリースしました。","category":"業界動向","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBMはApache 2.0とOpenMDW 1.0の両方のライセンスを使用するGranite Time Series PatchTST-FM-r2をリリースしました - Aioga AIニュース","description":"IBMは約3億8千5百万パラメータを持ち、最大8,192のコンテキスト、99分位数確率予測、欠損値補間に対応したGranite Time Series PatchTST-FM-r2をリリースしました。","url":"https://www.aioga.com/ja/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:47:55.912Z"},"ko":{"title":"IBM은 Apache 2.0과 OpenMDW 1.0 라이선스를 모두 사용하는 Granite Time Series PatchTST-FM-r2를 출시했습니다","summary":"IBM은 약 3억 8천 8천 개의 매개변수를 가진 Granite Time Series PatchTST-FM-r2를 출시했으며, 최대 8,192개의 문맥, 99분위수 확률 예측, 누락값 보간을 지원합니다.","category":"업계 동향","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM은 Apache 2.0과 OpenMDW 1.0 라이선스를 모두 사용하는 Granite Time Series PatchTST-FM-r2를 출시했습니다 - Aioga AI 뉴스","description":"IBM은 약 3억 8천 8천 개의 매개변수를 가진 Granite Time Series PatchTST-FM-r2를 출시했으며, 최대 8,192개의 문맥, 99분위수 확률 예측, 누락값 보간을 지원합니다.","url":"https://www.aioga.com/ko/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:04.525Z"},"es":{"title":"IBM lanzó Granite Time Series PatchTST-FM-r2, que utiliza licencias tanto de Apache 2.0 como de OpenMDW 1.0","summary":"IBM lanzó Granite Time Series PatchTST-FM-r2, con aproximadamente 385M de parámetros, soportando hasta 8.192 contextos, predicción de probabilidad de 99 cuantiles e interpolación de valores ausentes.","category":"Industria","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM lanzó Granite Time Series PatchTST-FM-r2, que utiliza licencias tanto de Apache 2.0 como de OpenMDW 1.0 - Aioga Noticias de IA","description":"IBM lanzó Granite Time Series PatchTST-FM-r2, con aproximadamente 385M de parámetros, soportando hasta 8.192 contextos, predicción de probabilidad de 99 cuantiles e interpolación d...","url":"https://www.aioga.com/es/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:04.381Z"},"fr":{"title":"IBM a publié Granite Time Series PatchTST-FM-r2, qui utilise à la fois les licences Apache 2.0 et OpenMDW 1.0","summary":"IBM a publié Granite Time Series PatchTST-FM-r2, avec environ 385 M de paramètres, prenant en charge jusqu’à 8 192 contextes, une prédiction de probabilité à 99 quantiles et une interpolation de valeurs manquantes.","category":"Industrie","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM a publié Granite Time Series PatchTST-FM-r2, qui utilise à la fois les licences Apache 2.0 et OpenMDW 1.0 - Aioga Actualités IA","description":"IBM a publié Granite Time Series PatchTST-FM-r2, avec environ 385 M de paramètres, prenant en charge jusqu’à 8 192 contextes, une prédiction de probabilité à 99 quantiles et une in...","url":"https://www.aioga.com/fr/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:13.378Z"},"de":{"title":"IBM veröffentlichte die Granite Time Series PatchTST-FM-r2, die sowohl Apache 2.0- als auch OpenMDW 1.0-Lizenzen verwendet","summary":"IBM veröffentlichte Granite Time Series PatchTST-FM-r2 mit etwa 385 Millionen Parametern, die bis zu 8.192 Kontexte, 99-Quantil-Wahrscheinlichkeitsvorhersage und Interpolation mit fehlenden Werten unterstützen.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM veröffentlichte die Granite Time Series PatchTST-FM-r2, die sowohl Apache 2.0- als auch OpenMDW 1.0-Lizenzen verwendet - Aioga KI-News","description":"IBM veröffentlichte Granite Time Series PatchTST-FM-r2 mit etwa 385 Millionen Parametern, die bis zu 8.192 Kontexte, 99-Quantil-Wahrscheinlichkeitsvorhersage und Interpolation mit...","url":"https://www.aioga.com/de/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:14.274Z"},"pt-BR":{"title":"A IBM lançou o Granite Time Series PatchTST-FM-r2, que utiliza licenças Apache 2.0 e OpenMDW 1.0","summary":"A IBM lançou o Granite Time Series PatchTST-FM-r2, com aproximadamente 385M de parâmetros, suportando até 8.192 contextos, previsão de probabilidade de 99 quantis e interpolação de valores ausentes.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"A IBM lançou o Granite Time Series PatchTST-FM-r2, que utiliza licenças Apache 2.0 e OpenMDW 1.0 - Aioga Notícias de IA","description":"A IBM lançou o Granite Time Series PatchTST-FM-r2, com aproximadamente 385M de parâmetros, suportando até 8.192 contextos, previsão de probabilidade de 99 quantis e interpolação de...","url":"https://www.aioga.com/pt-BR/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:23.388Z"},"ru":{"title":"IBM выпустила Granite Time Series PatchTST-FM-r2, использующий лицензии Apache 2.0 и OpenMDW 1.0","summary":"IBM выпустила Granite Time Series PatchTST-FM-r2 с примерно 385 млн параметров, поддерживающей до 8 192 контекста, предсказание вероятности на 99 квантилей и интерполяцию отсутствующих значений.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM выпустила Granite Time Series PatchTST-FM-r2, использующий лицензии Apache 2.0 и OpenMDW 1.0 - Aioga Новости ИИ","description":"IBM выпустила Granite Time Series PatchTST-FM-r2 с примерно 385 млн параметров, поддерживающей до 8 192 контекста, предсказание вероятности на 99 квантилей и интерполяцию отсутству...","url":"https://www.aioga.com/ru/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:23.269Z"},"ar":{"title":"أصدرت IBM سلسلة Granite Time Series PatchTST-FM-r2، التي تستخدم تراخيص Apache 2.0 وOpenMDW 1.0","summary":"أصدرت IBM سلسلة Granite Time Series PatchTST-FM-r2، بحوالي 385 مليون معلمة، تدعم حتى 8,192 سياقا، وتنبؤ احتمالية ب 99 كميا، واستيفاء القيم المفقودة.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"أصدرت IBM سلسلة Granite Time Series PatchTST-FM-r2، التي تستخدم تراخيص Apache 2.0 وOpenMDW 1.0 - Aioga أخبار الذكاء الاصطناعي","description":"أصدرت IBM سلسلة Granite Time Series PatchTST-FM-r2، بحوالي 385 مليون معلمة، تدعم حتى 8,192 سياقا، وتنبؤ احتمالية ب 99 كميا، واستيفاء القيم المفقودة.","url":"https://www.aioga.com/ar/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:31.922Z"},"hi":{"title":"आईबीएम ने ग्रेनाइट टाइम सीरीज़ पैचटीएसटी-एफएम-आर2 जारी किया, जो अपाचे 2.0 और ओपनएमडीडब्ल्यू 1.0 लाइसेंस दोनों का उपयोग करता है","summary":"आईबीएम ने लगभग 385M मापदंडों के साथ ग्रेनाइट टाइम सीरीज़ पैचटीएसटी-एफएम-आर2 जारी किया, जो 8,192 संदर्भों, 99-मात्रात्मक संभाव्यता भविष्यवाणी और लापता मूल्य प्रक्षेप का समर्थन करता है।","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"आईबीएम ने ग्रेनाइट टाइम सीरीज़ पैचटीएसटी-एफएम-आर2 जारी किया, जो अपाचे 2.0 और ओपनएमडीडब्ल्यू 1.0 लाइसेंस दोनों का उपयोग करता है - Aioga AI समाचार","description":"आईबीएम ने लगभग 385M मापदंडों के साथ ग्रेनाइट टाइम सीरीज़ पैचटीएसटी-एफएम-आर2 जारी किया, जो 8,192 संदर्भों, 99-मात्रात्मक संभाव्यता भविष्यवाणी और लापता मूल्य प्रक्षेप का समर्थन करता...","url":"https://www.aioga.com/hi/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:32.283Z"},"it":{"title":"IBM ha rilasciato Granite Time Series PatchTST-FM-r2, che utilizza sia le licenze Apache 2.0 che OpenMDW 1.0","summary":"IBM ha rilasciato la Granite Time Series PatchTST-FM-r2, con circa 385 M parametri, supportando fino a 8.192 contesti, previsione di probabilità a 99 quanti e interpolazione dei valori mancanti.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM ha rilasciato Granite Time Series PatchTST-FM-r2, che utilizza sia le licenze Apache 2.0 che OpenMDW 1.0 - Aioga Notizie IA","description":"IBM ha rilasciato la Granite Time Series PatchTST-FM-r2, con circa 385 M parametri, supportando fino a 8.192 contesti, previsione di probabilità a 99 quanti e interpolazione dei va...","url":"https://www.aioga.com/it/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:41.344Z"},"nl":{"title":"IBM bracht Granite Time Series PatchTST-FM-r2 uit, die zowel Apache 2.0- als OpenMDW 1.0-licenties gebruikt","summary":"IBM bracht Granite Time Series PatchTST-FM-r2 uit, met ongeveer 385 miljoen parameters, die tot 8.192 contexten, 99-quantiel kansvoorspelling en ontbrekende waarde interpolatie ondersteunt.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM bracht Granite Time Series PatchTST-FM-r2 uit, die zowel Apache 2.0- als OpenMDW 1.0-licenties gebruikt - Aioga AI-nieuws","description":"IBM bracht Granite Time Series PatchTST-FM-r2 uit, met ongeveer 385 miljoen parameters, die tot 8.192 contexten, 99-quantiel kansvoorspelling en ontbrekende waarde interpolatie ond...","url":"https://www.aioga.com/nl/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:40.879Z"},"tr":{"title":"IBM, hem Apache 2.0 hem de OpenMDW 1.0 lisanslarını kullanan Granite Zaman Serisi PatchTST-FM-r2'yi yayımladı","summary":"IBM, yaklaşık 385M parametreye sahip Granite Time Series PatchTST-FM-r2'yi yayımladı; bu parametre 8.192'ye kadar bağlamı, 99 quantile olasılık tahminini ve eksik değer interpolasyonunu destekledi.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM, hem Apache 2.0 hem de OpenMDW 1.0 lisanslarını kullanan Granite Zaman Serisi PatchTST-FM-r2'yi yayımladı - Aioga AI Haberleri","description":"IBM, yaklaşık 385M parametreye sahip Granite Time Series PatchTST-FM-r2'yi yayımladı; bu parametre 8.192'ye kadar bağlamı, 99 quantile olasılık tahminini ve eksik değer interpolasy...","url":"https://www.aioga.com/tr/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:49.921Z"},"vi":{"title":"IBM đã phát hành bản vá Granite Time Series TST-FM-r2, sử dụng cả giấy phép Apache 2.0 và OpenMDW 1.0","summary":"IBM đã phát hành Granite Time Series PatchTST-FM-r2, với khoảng 385 triệu tham số, hỗ trợ tối đa 8.192 ngữ cảnh, dự đoán xác suất 99 lượng tử và nội suy giá trị còn thiếu.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM đã phát hành bản vá Granite Time Series TST-FM-r2, sử dụng cả giấy phép Apache 2.0 và OpenMDW 1.0 - Tin tức AI Aioga","description":"IBM đã phát hành Granite Time Series PatchTST-FM-r2, với khoảng 385 triệu tham số, hỗ trợ tối đa 8.192 ngữ cảnh, dự đoán xác suất 99 lượng tử và nội suy giá trị còn thiếu.","url":"https://www.aioga.com/vi/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:50.469Z"},"id":{"title":"IBM merilis Granite Time Series PatchTST-FM-r2, yang menggunakan lisensi Apache 2.0 dan OpenMDW 1.0","summary":"IBM merilis Granite Time Series PatchTST-FM-r2, dengan sekitar 385M parameter, mendukung hingga 8.192 konteks, prediksi probabilitas 99 kuantil, dan interpolasi nilai hilang.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM merilis Granite Time Series PatchTST-FM-r2, yang menggunakan lisensi Apache 2.0 dan OpenMDW 1.0 - Berita AI Aioga","description":"IBM merilis Granite Time Series PatchTST-FM-r2, dengan sekitar 385M parameter, mendukung hingga 8.192 konteks, prediksi probabilitas 99 kuantil, dan interpolasi nilai hilang.","url":"https://www.aioga.com/id/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:59.601Z"},"th":{"title":"IBM ได้ปล่อย Granite Time Series PatchTST-FM-r2 ซึ่งใช้ทั้งใบอนุญาต Apache 2.0 และ OpenMDW 1.0","summary":"IBM ได้เปิดตัว Granite Time Series PatchTST-FM-r2 ที่มีพารามิเตอร์ประมาณ 385 ล้านตัว รองรับบริบทได้สูงสุด 8,192 บริบท การทํานายความน่าจะเป็น 99 ควอนไทล์ และการประมาณค่าที่ขาดหายไป","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM ได้ปล่อย Granite Time Series PatchTST-FM-r2 ซึ่งใช้ทั้งใบอนุญาต Apache 2.0 และ OpenMDW 1.0 - ข่าว AI Aioga","description":"IBM ได้เปิดตัว Granite Time Series PatchTST-FM-r2 ที่มีพารามิเตอร์ประมาณ 385 ล้านตัว รองรับบริบทได้สูงสุด 8,192 บริบท การทํานายความน่าจะเป็น 99 ควอนไทล์ และการประมาณค่าที่ขาดหายไป","url":"https://www.aioga.com/th/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:48:59.645Z"},"pl":{"title":"IBM wydał Granite Time Series PatchTST-FM-r2, który korzysta zarówno z licencji Apache 2.0, jak i OpenMDW 1.0","summary":"IBM wydał Granite Time Series PatchTST-FM-r2, z około 385 milionami parametrów, obsługującym do 8 192 kontekstów, przewidywanie prawdopodobieństwa 99-kwantowego oraz interpolację wartości brakujących.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"IBM wydał Granite Time Series PatchTST-FM-r2, który korzysta zarówno z licencji Apache 2.0, jak i OpenMDW 1.0 - Aioga Wiadomości AI","description":"IBM wydał Granite Time Series PatchTST-FM-r2, z około 385 milionami parametrów, obsługującym do 8 192 kontekstów, przewidywanie prawdopodobieństwa 99-kwantowego oraz interpolację w...","url":"https://www.aioga.com/pl/news/cmtuaa1fw14xzrofpxqrrz1t3/","contentTranslated":true,"sourceHash":"b21fcc7fbefe9444","translatedAt":"2026-09-09T16:49:08.258Z"}},"evidenceTier":"verified-news","reviewStatus":"automated-ingest","indexable":true,"editorialCover":""}}