IBM Releases Open-Source Time Series Model PatchTST-FM-r2

IBM Research has released its latest time series foundation model, Granite Time Series PatchTST-FM-r2. This model, with approximately 385M parameters, adopts a dual license of Apache 2.0 and OpenMDW 1.0, facilitating commercial use.

Architecture and Performance

The internal structure was redesigned compared to the previous r1 version to enhance the ability to capture long- and short-term temporal structures. Instead of a standard Transformer, Conformer blocks (Multi-head self-attention + Temporal convolution) were introduced to handle short-term patterns via convolution and long-term dependencies via attention. It supports a context of up to 8,192 steps and provides probabilistic forecasting based on 99 quantiles along with missing value interpolation.

In the GIFT-Eval benchmark, it achieved CRPS 0.467 and MASE 0.6846 among reproducible zero-shot models, ranking 1st among open-source licensed models. Despite its smaller size compared to some pre-trained large models, it demonstrated superior performance over competitors such as Chronos-2 and Timer-S1.

Deployment and Usage

Model weights, architecture, and inference pipeline code are all publicly available and can be easily loaded from the Hugging Face Hub via the granite-tsfm library. Without fine-tuning, users can input recent time series history to immediately generate forecasts and uncertainty intervals. IBM also offers an Early Access program in collaboration with Confluent Cloud, allowing direct inference of the model in Apache Flink-based streaming applications.