IBM's New AI Forecasting Model Can Predict Demand, Energy and Traffic Without Any Training on Your Data

Key points

• IBM released Granite Time Series PatchTST-FM-r2 on September 8, 2026, a roughly 385-million-parameter AI model built for forecasting.
• As of that date, it ranked first among zero-shot forecasting models with a permissive open-source licence on the GIFT-Eval benchmark, a standardised forecasting leaderboard.
• The model is free to use commercially under the Apache 2.0 licence, meaning businesses can build products with it without paying IBM.
• It handles up to 8,192 historical data points as context and returns 99 levels of probability estimates, not just a single prediction.
• IBM published the model weights, architecture and the code needed to reproduce its benchmark scores on Hugging Face.

IBM has released a new AI forecasting model, and the headline feature is not the performance score. It is the price: free, for commercial use, with no strings attached.

The model is called Granite Time Series PatchTST-FM-r2. Strip away the branding and it is a piece of software that looks at a stream of numbers over time, whether that is weekly sales figures, hourly electricity demand or network traffic, and predicts what comes next. The technical term for this kind of data is a time series, a sequence of measurements taken at regular intervals.

What makes this different from existing forecasting tools?

Most forecasting software requires you to train a separate model on your own historical data before it can predict anything. That takes time, expertise and computing costs. This model works zero-shot, meaning it can make useful predictions on data it has never seen before, with no training required on your end.

That matters most for smaller businesses or teams without data scientists on staff. You point the model at your data, and it starts working.

On GIFT-Eval, a standardised benchmark that tests forecasting models across many different types of data, PatchTST-FM-r2 ranked second overall among zero-shot models. Among models with a permissive commercial licence, it ranked first. GIFT-Eval scores models on two measures: CRPS and MASE, both of which reward accuracy and reward lower numbers. PatchTST-FM-r2 scored 0.467 on CRPS and 0.685 on MASE.

A note on survivorship bias here: leaderboard rankings measure performance on benchmark datasets, not on your specific business problem. A model that tops a leaderboard may still need tuning before it handles the quirks of real operational data reliably.

What changed under the hood?

The architecture, the underlying structure of the model, was redesigned from the previous version. IBM replaced standard transformer layers, the attention-based building blocks common in modern AI, with conformer layers, a design originally developed for speech recognition that combines attention with a convolution step, a filter that picks up short-term local patterns. The result is a model that spots both long-range trends and short-term fluctuations at the same time.

IBM also published exactly what data the model was trained on, including synthetic sequences generated inside the lab and public datasets kept separate from the benchmark test sets. For businesses worried about data contamination, that transparency is useful.

Should businesses actually try this?

If your organisation currently spends money on demand planning software or pays data engineers to maintain forecasting pipelines, this model is worth a serious look. It is free, the weights and code are public on Hugging Face, and IBM documented its training data clearly.

The honest takeaway: download the model, run it against a slice of your own historical data, and compare the output to whatever you use today. That test costs you an afternoon, not a procurement cycle.