IBM Releases Open Zero-Shot Model for Forecasting Time-Series Data

Teams forecasting demand, energy loads, prices or machine telemetry can now download an IBM model built to make predictions on unfamiliar data without retraining it for each task. IBM has released Granite PatchTST-FM-r2, a roughly 385-million-parameter time-series model that also produces uncertainty estimates and can fill in missing values.

IBM published the model’s weights, architecture, inference pipeline and code for reproducing its benchmark results. It is dual-licensed under Apache 2.0 and OpenMDW 1.0, giving users a choice of two permissive licenses for the released materials. The model is available on the Hugging Face Hub and loads through version 0.3.9 or later of the granite-tsfm package.

Time-series forecasting works with ordered measurements: sales by day, electricity use by hour or sensor readings over time. PatchTST-FM-r2 is designed for zero-shot forecasting, meaning it generates a forecast on new data without fine-tuning or task-specific fitting. IBM says it supports probabilistic forecasting, producing point forecasts alongside intervals that express uncertainty, and missing-value imputation.

The release retains the patch-based approach of its predecessor but swaps standard transformer blocks for conformer blocks. Those blocks combine attention, which can model longer relationships among chunks of data, with temporal convolution intended to capture short-range patterns. IBM says the split lets attention concentrate on longer-range relationships between patches.

What the model is built to return

• Forecasts across flexible future lengths from input histories of up to 8,192 steps.
• Ninety-nine quantiles, which can be used to express a range of plausible future values rather than a single number.
• Filled-in values where a time series contains gaps.

IBM reported that PatchTST-FM-r2 ranked second among replicable zero-shot models on GIFT-Eval for both CRPS and MASE. But the model card says the results are in a pending pull request to the benchmark, and dates that standing to August 31; IBM’s announcement dates the same position to September 8. The published reproduction materials make the claim more inspectable, but the release does not establish how the model will perform on a particular organization’s own data.

The model is the successor to PatchTST-FM-r1 and is positioned for regularly sampled series such as demand, prices, energy loads, traffic and telemetry. Its open release gives developers a route to test the model’s forecasts, uncertainty intervals and handling of missing data against the patterns that matter in their own operations. That local comparison—not a general benchmark rank alone—will determine whether zero-shot convenience translates into a useful forecasting workflow.