IBM Releases 385M-Parameter Granite Time Series Forecasting Model

IBM released the Granite Time Series PatchTST-FM-r2, a 385-million parameter time-series foundation model for zero-shot forecasting. [1] The model supports forecasting tasks for demand, prices, energy loads, traffic, and telemetry data. [1] The system features an updated architecture, a larger pretraining corpus, support for missing value imputation, and a 99-quantile prediction head for probabilistic forecasting. [1] The model provides a context length of up to 8,192 with flexible forecast lengths, and its backbone is constructed using conformer blocks. [1]

As of September 8, 2026, PatchTST-FM-r2 ranks second overall among replicable, zero-shot models on the GIFT-Eval time series forecasting benchmark. [1] It is the top-performing model in that category released under permissive, commercial-friendly licenses, specifically Apache 2.0 and OpenMDW 1.0. [1] IBM has made the model weights, architecture, inference pipeline, and code required to reproduce benchmark results available on Hugging Face. [1] Additionally, models from the Granite Time Series family can be integrated into production streaming applications using Confluent products. [1]

The release provides enterprise developers with a highly capable forecasting tool that avoids the need to train separate models for individual datasets. By securing the top spot for commercial-friendly licenses on the GIFT-Eval benchmark, IBM positions PatchTST-FM-r2 as a viable open-source alternative for production environments, particularly for teams already using Confluent for streaming data. What the sources don't address: how the model's latency scales in live streaming environments compared to smaller, task-specific forecasting models.