IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license
| Source: Hugging Face Blog
Tags: IBM, Granite, time series, forecasting, Apache 2.0, GIFT-Eval, zero-shot
IBM's Granite Time Series PatchTST-FM-r2 takes the top spot among replicable zero-shot commercial models on the GIFT-Eval benchmark — a 385M-parameter model with 8,192 context length, probabilistic forecasting, and Apache 2.0 licensing available now on Hugging Face.
Details
IBM Research released Granite Time Series PatchTST-FM-r2, the latest in its TSFM (Time Series Foundation Model) family, on September 8, 2026. The model tops the GIFT-Eval leaderboard for replicable, zero-shot models with commercial-friendly licenses, and ranks second overall among all replicable zero-shot models on the same benchmark. The model offers approximately 385M parameters, supports context lengths up to 8,192, and delivers probabilistic forecasts via a 99-quantile prediction head. Its backbone combines conformer blocks with multi-head self-attention and temporal convolution — designed to capture both long-range and short-range temporal dependencies. Target use cases include demand forecasting, energy loads, traffic, prices, and telemetry. What distinguishes this release from similar announcements is the licensing: dual Apache 2.0 and OpenMDW 1.0, meaning users can pick either and commercial deployment is straightforward. IBM has published weights, architecture, inference pipeline, and code sufficient to reproduce the benchmark results. The blog post also notes integration with Confluent for streaming production deployments. For organizations currently maintaining per-domain forecasting models, this is a credible zero-shot baseline worth evaluating against.