IBM has announced the release of the Granite Time Series PatchTST-FM-r2, a significant advancement in time-series forecasting technology. This model, part of the Granite TSFM family, is designed to revolutionize how forecasting systems operate by enabling users to generate forecasts without the need for extensive retraining on specific datasets.
Overview of PatchTST-FM-r2
PatchTST-FM-r2 builds upon its predecessor, PatchTST-FM-r1, by integrating an updated architecture and a larger pretraining corpus. This model boasts approximately 385 million parameters and offers robust zero-shot performance, making it the top-performing model in its category as of September 8, 2026. It is dual-licensed under Apache 2.0 and OpenMDW 1.0, providing a permissive framework for commercial use.
Performance Metrics
In the GIFT-Eval benchmark, which evaluates forecasting models across various scenarios, PatchTST-FM-r2 ranks second among replicable, zero-shot models for both CRPS and MASE metrics. Specifically, it achieves a geometric-mean CRPS of 0.467 and a geometric-mean MASE of 0.6846. Notably, it is the highest-performing model with a permissive license in this category.
Architectural Innovations
The architecture of PatchTST-FM-r2 retains the effective patch-based representation of its predecessor while introducing conformer layers that combine multi-head self-attention with temporal convolution. This design enhances the model’s ability to capture both long- and short-term temporal relationships, improving forecasting accuracy. The model supports context lengths of up to 8,192 steps and provides probabilistic forecasts through a 99-quantile prediction head.
Training Data Transparency
IBM emphasizes the importance of transparency in the training data used for the model. PatchTST-FM-r2 is trained on a documented corpus that includes selected datasets from GIFT-Eval, custom synthetic data, and approximately 500,000 synthetic CauKer sequences. This clarity aids organizations in assessing the model’s applicability and governance.
With the release of PatchTST-FM-r2, IBM aims to empower researchers and enterprises to leverage advanced forecasting capabilities in various applications, from demand and energy load forecasting to traffic and telemetry analysis. The model is readily accessible for experimentation and integration into production environments, including streaming applications through partnerships with platforms like Confluent.
This article was produced by NeonPulse.today using human and AI-assisted editorial processes, based on publicly available information. Content may be edited for clarity and style.
