On September 9, IBM detailed Granite Time Series PatchTST-FM-r2. With about 385 million parameters, it forecasts numerical sequences without additional training for the particular task. Weights and code are available; users can choose Apache 2.0 or OpenMDW 1.0 for the model.

IBM's September comparison places it second among replicable zero-shot models on GIFT-Eval. This is a defined category and a manufacturer assessment. The ranking alone does not establish how well it will handle a particular company's irregular data.

Our editorial perspective: a building manager may find a range for tomorrow's energy demand more useful than one seemingly precise number. An open model makes it possible to test this on an organization's own server and compare it with a simple seasonal forecast.

A meaningful pilot must separate the history available at forecast time from later outcomes. Alongside average error, it should examine missing measurements, unusual events, costs and whether uncertainty intervals match what actually happens.

Our optimistic editorial estimate is 2–6 weeks for a team with prepared data to evaluate a limited internal pilot. Automated decision-making requires further operational validation; this interval is not a manufacturer promise.