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

PatchTST-FM-r2 is the latest model in IBM's Granite TSFM family and a new version of PatchTST-FM-r1. According to the blog post, it combines an updated architecture, a larger pretraining corpus, probabilistic forecasting, support for imputation of missing values, and strong zero-shot performance in a model of about 385 million parameters. The model is presented as a general-purpose tool for zero-shot forecasting of demand, prices, energy loads, traffic, telemetry, and other time series. It has a context length of up to 8,192, flexible forecast lengths, and probabilistic forecasts produced through a 99-quantile prediction head.

GIFT-Eval is described as a comprehensive time-series forecasting benchmark designed to evaluate models across diverse forecasting scenarios. When the leaderboard is restricted to models that are zero-shot, replicable, and evaluated without test leakage, PatchTST-FM-r2 ranks second for both CRPS and MASE as of Sept. 8, 2026, with lower values being better for both metrics. The blog post says the model achieves a geometric-mean CRPS of 0.467, placing it immediately behind TimesFM-3 in that comparison and first among models with permissive licenses. Its geometric-mean MASE is 0.6846.

Some models on GIFT-Eval are categorized as pretrained rather than strictly zero-shot, and those models are allowed to include the training portions of GIFT-Eval evaluation datasets in their pretraining corpora. Even when pretrained models are added to the comparison, PatchTST-FM-r2 remains near the top, ranking third for CRPS and fourth for MASE among replicable models, according to the blog post. It outperforms several pretrained models, including Chronos-2, Timer-S1, and Toto variants, even though some competing models are considerably larger.

On architecture, PatchTST-FM-r2 retains the patch-based representation that made the PatchTST family effective, but its internal design is changed to capture long- and short-term relationships efficiently and to smoothen inter-patch predictions. The blog post says both changes substantially improve error measures. One change is the move from standard transformer layers to layers that incorporate convolution along with multi-head self-attention. These are called conformer layers and originate in speech processing applications. A PatchTST-FM-r1 block combined multi-head self-attention with a feed-forward network; in r2, that block was replaced with a conformer-style block.

The model weights, architecture, inference pipeline, and code needed to reproduce the benchmark results are available, according to the blog post. The blog post also highlights how models from the Granite Time Series family can be used in streaming applications in production settings by leveraging a Confluent product.