On September 2, 2026, IBM Research published the Granite Time Series PatchTST-FM-r2 model on Hugging Face. The 0.4-billion-parameter model is released under the Apache 2.0 license, making it available for commercial use without restriction.

The model card on Hugging Face notes architectural enhancements over its predecessor and an expanded training base. The update is dated August 31, 2026. IBM's research blog confirms the Granite time-series family now includes models for every major task — forecasting, anomaly detection, and classification.

The core TSFM models are accessible through the Hugging Face transformers library. The granite-tsfm GitHub repository provides notebooks, utilities, and serving components for Python 3.10 through 3.13. The repository also includes Google Colab tutorials for TinyTimeMixer and demos from NeurIPS 2023.

PatchTST-FM-r2 builds on the March 2026 release of four time-series foundation models: FlowState r1.1, TTM-r3, PatchTST-FM r1, and TSPulse r1. The predecessor PatchTST-FM-r1 carries 0.3B parameters and was updated on March 25, 2026.

Confirmed

  • Model name: ibm-granite/granite-timeseries-patchtst-fm-r2 (PatchTST-FM-r2)
  • Parameter count: 0.4B
  • License: Apache 2.0 (commercial-friendly)
  • Release date on Hugging Face: August 31, 2026 (shown as "9 days ago" from the blog publication date of September 2, 2026)
  • Tasks supported: forecasting, anomaly detection, classification
  • Available via: Hugging Face transformers, granite-tsfm GitHub repo, and IBM Granite documentation
  • Predecessor: PatchTST-FM-r1 (0.3B, released March 25, 2026)
  • March 2026 quad release: FlowState r1.1, TTM-r3, PatchTST-FM r1, TSPulse r1
  • GitHub repo supports Python 3.10, 3.11, 3.12, 3.13
  • Notebooks include: PatchTSMixer, PatchTST, TinyTimeMixer, FlowState
  • Google Colab tutorial available for TTM

Unknown

  • Exact training dataset composition and size for the expanded training base
  • Independent benchmark results against competing time-series foundation models (e.g., Chronos, TimesFM, Moirai)
  • Inference latency and memory footprint on CPU-only vs. GPU deployments
  • Whether the Apache 2.0 license applies to model weights, code, and documentation uniformly
  • Long-term maintenance and update cadence for the r2 series

Our take

IBM's move to Apache 2.0 for a state-of-the-art time-series foundation model is a deliberate play for enterprise adoption where licensing friction often blocks open-weight models. The 0.4B parameter count keeps inference light enough for CPU-only deployment — a practical advantage for regulated environments that cannot ship data to GPU clouds. What remains unproven is whether the expanded training base translates to measurable gains on proprietary enterprise datasets without fine-tuning.

Sources