IBM has released Granite Time Series PatchTST-FM-r2, a 385-million-parameter foundation model designed for zero-shot time-series forecasting across applications including demand prediction, pricing, energy loads, and traffic monitoring. Published on Hugging Face with dual Apache 2.0 and OpenMDW 1.0 licenses, the model ranks as the top-performing option in the commercial-friendly category on the GIFT-Eval benchmark as of September 8, 2026, achieving competitive results even against significantly larger proprietary models.
The updated model introduces several architectural improvements over its predecessor, most notably the adoption of conformer-based layers that combine multi-head self-attention with temporal convolution to more effectively capture both long-term and short-term patterns. The system supports up to 8,192-token context lengths, probabilistic forecasting through 99-quantile predictions, and missing-value imputation, making it production-ready for enterprise applications. IBM has published the model weights, inference pipeline, and code necessary to reproduce benchmark results.
The permissive licensing and strong zero-shot generalization position the model as an open-source alternative to proprietary forecasting solutions. By releasing the full implementation publicly, IBM aims to democratize access to foundation model technology for time-series applications across industries.
Key Points
IBM released PatchTST-FM-r2, a 385M-parameter model under permissive open licenses (Apache 2.0 and OpenMDW 1.0)
Model ranks #1 among commercially-licensed zero-shot options and #2 overall on GIFT-Eval forecasting benchmark
Architecture upgraded with conformer blocks integrating self-attention and temporal convolution for improved pattern recognition
Complete implementation open-sourced including weights, inference pipeline, and benchmark reproduction code for production deployment