IBM Releases New Open Model for Forecasting Data Over Time

IBM has released Granite Time Series PatchTST-FM-r2, the newest model in its Granite time-series foundation-model family, aimed at forecasting things like demand, prices, energy loads, traffic and other data that changes over time. As of September 8, 2026, the roughly 385-million-parameter model is the top-performing "zero-shot" model released under a permissive, commercial-friendly open-source license on the GIFT-Eval leaderboard, a benchmark that tests forecasting models across a wide range of scenarios.

A zero-shot model can generate forecasts for a new dataset without being separately trained on it first, unlike traditional forecasting systems that need a dedicated model for every dataset. PatchTST-FM-r2 can handle input histories of up to 8,192 time steps, produce forecasts of flexible length, and give probabilistic forecasts, meaning it estimates a range of likely outcomes rather than a single number.

On GIFT-Eval, restricted to zero-shot models that are replicable and evaluated without test leakage, PatchTST-FM-r2 ranks second overall, behind TimesFM-3, and first among models with permissive licenses such as its own dual Apache-2.0 and OpenMDW-1.0 terms. Even when the comparison is widened to include "pretrained" models that are allowed to train on parts of the benchmark's own datasets, PatchTST-FM-r2 still ranks near the top, outperforming several larger pretrained models including Chronos-2, Timer-S1 and Toto variants.

The new version changes the model's internal architecture. Its predecessor, PatchTST-FM-r1, combined self-attention, a technique that lets the model weigh relationships between different points in a time series, with a standard feed-forward network. PatchTST-FM-r2 instead uses "conformer" blocks, borrowed from speech-processing research, which pair self-attention with temporal convolution, a technique for capturing local patterns. Self-attention models long-range relationships in the data, while convolution captures shorter-term structure.

IBM says the model's weights, architecture, inference pipeline and code needed to reproduce the benchmark results are all publicly available. The company also highlighted how the Granite time-series models can be used in streaming applications through the Confluent data-streaming platform.