IBM launches Granite, an AI model for forecasting time series

IBM has released Granite Time Series PatchTST-FM-r2, an AI model that can forecast data it did not see during training. It could help anticipate demand, energy consumption, traffic, prices, sensor readings, or server usage without creating a separate model for each use case.

The key feature is zero-shot forecasting: you only need to provide the recent history of a time series, meaning a sequence of data points ordered by time, for it to generate a prediction. You do not need to fine-tune it beforehand with each company’s data.

An open model with strong results

According to IBM, the model ranks second among replicable zero-shot models on the GIFT-Eval benchmark, using data published on September 8, 2026. In this comparison, lower scores indicate better performance.

• It achieved a geometric mean CRPS of 0.467, a metric that evaluates the quality of predictions and their uncertainty.
• It achieved a MASE of 0.6846, which compares the model’s error with that of a reference method.
• It is the highest-ranked model in that category among those using permissive licenses suitable for commercial use.

When models that may have seen part of the evaluated datasets during training are also included, PatchTST-FM-r2 ranks third in CRPS and fourth in MASE. The result is relevant, but it does not mean the model is the best choice for every scenario or type of data.

The model has around 385 million parameters, supports histories of up to 8,192 steps, and can produce forecasts for different future lengths. It also generates up to 99 quantiles, allowing it to represent several possible scenarios and build uncertainty intervals instead of returning a single figure.

What changes from the previous version

IBM has redesigned part of the PatchTST-FM-r1 architecture. The new version combines two ways of analyzing a time series:

• Attention, which detects relationships between points that are far apart in time.
• Temporal convolutions, which capture local patterns and nearby changes.

This combination, known as a conformer architecture, can help the model distinguish between long-term trends and short-term behavior. It also uses overlapping segments of the series to smooth transitions between forecasts.

In addition, PatchTST-FM-r2 can fill in missing values and provides both point forecasts and distributions of possible outcomes. In practice, it could estimate not only that demand will reach 10,000 units, but also which range of values is reasonable.

Data and licensing

IBM has documented the sources used for training. They include selected GIFT-Eval datasets, synthetic data created with different methods, and around 500,000 synthetic sequences of 4,096 steps.

Transparency does not remove the need to evaluate the model before using it in a business, especially in regulated industries. It does make it easier to assess which data was used and whether there is a risk of contamination involving the datasets used for testing.

The weights, architecture, inference system, and code for reproducing the results are publicly available. The model is distributed under a dual license: Apache 2.0 and OpenMDW 1.0. Organizations can choose either license, and both provide broad permissions to use, modify, and distribute the model.

What it can do in practice

The development team can load the model from Hugging Face, provide the latest data for a metric, and request a forecast without training anything from scratch. That reduces the initial work for use cases such as:

• Forecasting sales or inventory.
• Estimating the load on an electric grid.
• Anticipating traffic or transaction volume.
• Detecting anomalous patterns in sensors and servers.
• Filling in missing sections of data.

It can also be used with data that arrives continuously. IBM and Confluent offer some Granite Time Series models through an early-access program within Confluent Cloud, where forecasting and anomaly detection run on real-time data streams using Apache Flink.

For you, the main difference is that this type of model reduces the need to build and maintain a separate solution for each time series. The next step is to check how it performs with each organization’s real data, because a strong benchmark result does not replace testing with the history, update rate, and errors specific to each business.