IBM has released the open-source model Granite Time Series PatchTST-FM-r2 for time series forecasting

IBM has released Granite Time Series PatchTST-FM-r2, a new version of the previous model PatchTST-FM-r1 from the Granite TSFM (time series foundation models) family. According to the company, as of 8 September 2026, it is the highest-rated zero-shot model with a permissive open-source license among reproducible models on the GIFT-Eval leaderboard, a benchmark for evaluating time series forecasting across different scenarios – it ranked 2nd overall among reproducible zero-shot models and 1st in the category of permissively licensed models (geometric mean CRPS 0.467, MASE 0.6846, according to the company).

The model has approximately 385 million parameters, a context window of up to 8192 and supports probabilistic forecasts through a head with 99 quantiles, flexible forecast lengths and filling in missing values (imputation). It is available under a dual license, Apache 2.0 and OpenMDW 1.0; users can choose which one to use, and both allow commercial use. The model weights, architecture, inference pipeline and code for reproducing the benchmark results are publicly available.

Architecturally, the model replaces standard transformer layers with so-called conformer blocks, which combine multi-head self-attention with temporal convolution – an approach borrowed from speech processing. According to the company, this combination allows the model to capture both short-term local patterns (through convolution) and long-term dependencies (through attention), which is also reflected in different attention patterns compared with the previous transformer version. The source article also describes the use of models from the Granite Time Series family in streaming applications with Confluent, but this section is not available in full in the source. Details can be found in the source article.