IBM releases Granite PatchTST-FM-r2, a top-ranked open zero-shot time-series forecasting model

IBM Research has released Granite Time Series PatchTST-FM-r2, a 385-million-parameter zero-shot forecasting model that IBM says is the top-ranked permissively-licensed model on the GIFT-Eval benchmark.

What's new

IBM published the model on Hugging Face under a dual license — "dual licensed under Apache 2.0 and OpenMDW 1.0" — letting users pick whichever permissive open-source terms fit their use case. The model has roughly 385 million parameters, a context length of up to 8,192 steps, and produces probabilistic forecasts through a 99-quantile prediction head rather than a single point estimate. Its backbone uses what IBM calls conformer blocks, combining multi-head self-attention with temporal convolution to "capture long- and short-range temporal structure" in a sequence.

On the GIFT-Eval leaderboard, a benchmark for time-series forecasting, PatchTST-FM-r2 ranks #2 overall among replicable, zero-shot models and #1 among permissively-licensed zero-shot models as of its September 8, 2026 release date, posting a geometric-mean CRPS of 0.467 and MASE of 0.6846. IBM notes it stays competitive even against models that were allowed to train on benchmark data directly, ranking 3rd on CRPS and 4th on MASE against that less restrictive comparison group. Like other zero-shot time-series foundation models, it forecasts on new data without any fine-tuning or task-specific fitting, and IBM published the weights, architecture, and inference pipeline needed to reproduce the benchmark results.

Context

Time-series forecasting — demand, prices, energy loads, traffic, sensor telemetry — is one of the less-visible corners of applied AI, dominated historically by statistical methods (ARIMA, exponential smoothing) and bespoke per-series models. Foundation-model-style zero-shot forecasters, trained once on a broad corpus of series and then applied to new domains without retraining, are a newer category that IBM's Granite Time Series line has been building out release over release, aiming for the same "pretrain once, deploy anywhere" economics that transformed language and vision AI.

The commercial licensing angle is deliberate positioning: many of the strongest zero-shot forecasters on leaderboards like GIFT-Eval carry research-only or restrictive licenses, or were trained in ways that make their benchmark standing hard to reproduce independently. IBM is explicitly targeting the gap between "best on the leaderboard" and "usable in a commercial product," publishing full reproduction materials alongside the weights.

Why it matters

This is a narrow but real enterprise use case — forecasting demand, inventory, energy load, and traffic is a recurring workload across manufacturing, retail, logistics, and utilities, and most of that work today still runs on older statistical tooling or in-house models that need per-series retraining. A permissively licensed, zero-shot model that is competitive with restrictively licensed leaderboard leaders lowers the bar for enterprises to adopt a foundation-model approach to forecasting without triggering the licensing reviews that block adoption of research-only checkpoints.

It also reinforces a broader pattern of foundation models pushing into specialized, previously bespoke domains — time series joins code, biology, and robotics as an area where a general pretraining approach is starting to challenge purpose-built statistical or domain-specific models on their own benchmarks.