IBM Granite Time Series Forecaster Goes Apache 2.0

A Forecasting Foundation Model With a Permissive License

IBM Research released Granite Time Series PatchTST-FM-r2 on September 9, and the headline is not the size of the model but the terms attached to it. At roughly 385 million parameters, this is a small model by 2026 standards — and it is dual-licensed under Apache 2.0 and OpenMDW 1.0, which means a company can put it into a commercial product without negotiating anything.

• Size and context: about 385M parameters, context lengths up to 8,192 steps, flexible forecast horizons
• Licensing: dual-licensed Apache 2.0 and OpenMDW 1.0 — pick whichever fits your legal posture
• GIFT-Eval standing (as of September 8, 2026): CRPS 0.467 and MASE 0.6846, second among replicable zero-shot models and first among those with permissive licenses
• Outputs: probabilistic forecasts via a 99-quantile prediction head, plus imputation of missing values

What Is a Time Series Foundation Model?

Most forecasting in production still works the old way: you fit a model to one series, for one horizon, and refit it when the data shifts. A time series foundation model is pretrained across a large corpus of unrelated series so it can forecast a new one zero-shot — no task-specific training run, no per-series tuning. Point it at demand, prices, energy loads, traffic or telemetry and ask for the next N steps.

That generalization is the whole value proposition for teams that own thousands of series and cannot afford a bespoke model for each. It is also why the benchmark matters. GIFT-Eval scores forecasters across a diverse pool of datasets, and lower is better on both metrics IBM reports: CRPS for the quality of the full probabilistic forecast, MASE for scaled point accuracy. Second place among replicable zero-shot models is a strong result for a model this small; first place among permissively licensed ones is the part that decides procurement meetings.

Inside the Architecture

The backbone is built from conformer blocks, which pair multi-head self-attention with temporal convolution so the model can pick up both long-range structure and the short, local wiggles that attention alone tends to smooth over. IBM also moved to 50% overlapping patches with Hamming-window weighting — a signal-processing touch that softens the boundaries between patches instead of chopping the series into hard segments.

The 99-quantile head deserves its own mention. A single predicted number is rarely what an operations team needs; they need a band. Getting quantiles directly from the model means uncertainty intervals come free rather than from a bolted-on wrapper.

How Do You Actually Run It?

IBM ships the model through Hugging Face Hub alongside its granite-tsfm package, and the documented path is short: install the package, load the checkpoint, and run historical data through the time series forecasting pipeline to get quantile outputs back. A getting-started notebook lives in the ibm-granite/granite-tsfm repository on GitHub.

The practical read: at 385M parameters this runs comfortably on modest hardware, including a single mid-range GPU or a capable CPU box, which puts it in reach of the same self-hosted setups people already use for local language models. If you have been watching open-weight releases, this is a different corner of the same trend — see our coverage of IBM Granite 4.2's open reasoning models and DeepSeek V4.1-Flash's MIT-licensed weights. More in our AI section.

Sources: IBM Research via Hugging Face — September 9, 2026; Unite.AI — September 2026.