IBM Releases Granite PatchTST-FM-r2 Zero-Shot Time Series Model Under Commercial License

On September 9, 2026, IBM uploaded a 3.07-gigabyte safetensors file to the Hugging Face Hub, releasing the Granite Time Series PatchTST-FM-r2 model. The 385-million-parameter system arrived with open weights, its inference pipeline, and reproducible benchmark code. Designed to predict future values across regularly sampled data, the model targets enterprise workflows like energy load balancing, retail demand, and machine telemetry.[1][4]

Time-series forecasting relies on ordered measurements—sales by day, electricity use by hour, or sensor readings over time. Historically, data science teams have had to train and maintain a separate machine learning model for every specific dataset. This fragmented approach creates massive operational overhead for companies tracking thousands of distinct metrics.[3]

"Time-series foundation models are changing the way forecasting systems are built," the IBM Research team wrote in their release announcement. "Instead of training and maintaining a separate model for every dataset, users can use a pretrained model and generate forecasts zero-shot."[1]

The model processes data through a patching mechanism. Instead of reading individual data points one by one, PatchTST-FM-r2 vectorizes time series into subseries-level patches. It uses 50 percent overlapping patches with Hamming-window weighting to smooth the boundaries between chunks of data, returning a continuous prediction.[1][6]

The internal architecture represents a significant shift from its predecessor, PatchTST-FM-r1. IBM replaced standard transformer blocks with 30 conformer blocks—a design originally developed for speech processing.[2][6]

Each conformer block surrounds a multi-head self-attention mechanism with a temporal convolution layer. The convolution component captures short-range temporal structures and local patterns, freeing the attention mechanism to concentrate on long-range relationships across a historical context window of up to 8,192 steps.[2][6]

For enterprise deployments, a single predicted value is rarely sufficient. PatchTST-FM-r2 produces probabilistic forecasts, outputting 99 distinct quantiles. This allows a supply chain planner to set reorder points based on the 90th percentile of expected demand, rather than relying on a rigid point forecast that ignores uncertainty.[3][4]

The model's release coincides with an integration into live data pipelines. IBM and Confluent launched an Early Access program on September 1, 2026, bringing Granite Time Series models directly into Confluent Cloud.[5]

Through Apache Flink SQL, teams can call the forecasting and anomaly-detection models on live data streams. Inference results are written back to Kafka topics, allowing the model to act on the current state of the business rather than stale batch data.[5]

On the GIFT-Eval benchmark, which evaluates models across diverse forecasting scenarios, PatchTST-FM-r2 achieved a geometric-mean Continuous Ranked Probability Score (CRPS) of 0.467 and a Mean Absolute Scaled Error (MASE) of 0.6846.[1][2]

As of September 8, 2026, those scores place the model second overall among replicable, zero-shot models, trailing only Google's TimesFM-3.[1][2]

The defining differentiator for PatchTST-FM-r2 is its licensing. While many open-weight models carry non-commercial restrictions, IBM dual-licensed the release under Apache 2.0 and the Linux Foundation's OpenMDW 1.0.[1][4]

"Time-series forecasting is one of the last ML domains where companies still hand-train per-dataset models — it's also a domain where 'open weights' often quietly means 'research only,'" noted an analysis by Compendia Labs. The permissive licensing removes the legal barriers to deploying the 385-million-parameter model in commercial, on-premises environments.[4]