IBM releases SOTA Granite Time Series model with commercial-friendly license (huggingface.co)

IBM has announced the release of its latest time-series foundation model, Granite Time Series PatchTST-FM-r2, designed to enhance forecasting capabilities across various applications. This model, which boasts approximately 385 million parameters, incorporates an innovative architecture that combines conformer blocks and traditional multi-head self-attention, enabling it to efficiently capture both long- and short-term relationships in data. PatchTST-FM-r2 excels in zero-shot performance, ranking as the top model under a permissive, commercial-friendly license on the GIFT-Eval leaderboard, making it accessible for organizations looking to leverage AI for forecasting without restrictive licensing barriers. The model’s technical improvements include support for probabilistic forecasting, imputation of missing values, and the ability to process lengthy input contexts of up to 8,192 steps. Its availability under dual licenses—Apache 2.0 and OpenMDW 1.0—provides developers and researchers with flexibility and confidence in its use. Importantly, the transparent pretraining process enables users to understand the data involved, thereby facilitating better model governance. With this release, IBM continues to strengthen its position in the AI/ML community, promoting wider adoption of advanced time-series forecasting solutions, especially beneficial for applications in demand prediction, energy consumption, and traffic management.