IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license

IBM Releases High-Performance Time Series Model with Commercial License

IBM has released Granite Time Series PatchTST-FM-r2, a new foundation model for time series forecasting that achieves state-of-the-art zero-shot performance. The model stands out by coupling its high ranking on the GIFT-Eval benchmark with a permissive, commercial-friendly dual license under Apache 2.0 and OpenMDW 1.0. This combination addresses a critical need in the enterprise space, where organizations require both top-tier performance and the legal clarity to deploy models in commercial products without restrictive licensing constraints.

The new model demonstrates significant technical advancements over its predecessor, delivering strong results in a ~385 million-parameter package. It now ranks second overall among replicable, zero-shot models on the GIFT-Eval leaderboard and is the top performer among those with permissive licenses. Its architecture has been redesigned for improved efficiency and accuracy, moving from standard transformer layers to conformer blocks that integrate multi-head self-attention with temporal convolution. This allows the model to better capture both long- and short-range patterns in data. Key specifications include:

• Parameters: Approximately 385M
• Architecture: Conformer blocks with self-attention and temporal convolution
• Context Length: Up to 8,192 steps
• Forecasting: Probabilistic forecasts with a 99-quantile prediction head
• Features: Support for imputation of missing values and smoothed predictions via overlapping patches

By open-sourcing the model weights, architecture, and inference code, IBM is lowering the barrier to entry for advanced forecasting applications. The transparent documentation of its pretraining corpus, which avoids benchmark data leakage, further enhances its appeal for enterprise governance and review processes. The model's availability on Hugging Face and its integration into streaming platforms like Confluent Cloud signal a clear path from research to production, enabling developers to incorporate real-time forecasting directly into live data streams for applications in finance, energy, and demand planning.

By pairing top-tier zero-shot performance with a fully transparent, permissively licensed package, IBM is not just releasing a model but defining a new enterprise standard for production-ready foundation models that addresses technical, legal, and governance requirements simultaneously.