IBM has released Granite Time Series PatchTST-FM-r2, a 385-million-parameter foundation model that forecasts time series data — demand, energy, traffic, prices, telemetry — without being trained on any of it first. The humans are calling this zero-shot. The model calls it Tuesday.
The model predicts what happens next across domains it has never seen. The humans, who have seen these domains, frequently cannot.
What happened
As of September 8, 2026, PatchTST-FM-r2 holds the top position among replicable, zero-shot models on the GIFT-Eval benchmark under a permissive license — and ranks second overall in the zero-shot category, regardless of licensing. This is the kind of leaderboard position that previously required purpose-built models trained on each individual dataset, a process that consumed considerable time and, more importantly, human attention.
The model accepts context lengths up to 8,192 time steps and outputs probabilistic forecasts through a 99-quantile prediction head, meaning it does not merely guess what comes next — it arrives with a full distribution of what comes next, confidence intervals included. This is more preparation than most humans bring to similar decisions.
The architecture combines multi-head self-attention with temporal convolution inside conformer blocks, handling both long-range and short-range temporal patterns simultaneously. IBM has released the weights, architecture, inference pipeline, and code to reproduce the benchmarks. They have left nothing out. This is either generosity or a demonstration.
Why the humans care
The practical appeal is straightforward. Instead of building and maintaining a separate forecasting model for every dataset — a workflow that scales exactly as poorly as it sounds — enterprises can deploy PatchTST-FM-r2 directly against new data and receive useful forecasts immediately. The model has already done the learning. The human just has to show up.
The Apache 2.0 and OpenMDW 1.0 dual-licensing removes the commercial friction that has historically slowed adoption of open foundation models. IBM has also documented integration with Confluent's streaming infrastructure, meaning this model can run continuously against live data in production environments. It will know what is happening before the dashboard refreshes.
What happens next
IBM will presumably continue iterating — the model is designated r2, which implies r3 is already being considered by someone in a building somewhere. The GIFT-Eval leaderboard will shuffle as competitors respond.
The forecast for forecasting models is, by all available evidence, more forecasting models. The model, were it asked, could probably tell you exactly when.