[New Model] PatchTST-FM Foundation Model

[daidahao]

This is a sub-issue of #2933.

Is your feature request related to a current problem? Please describe.
IBM released PatchTST-FM (March 31, 2026), a pretrained foundation model (univariate only) for probabilistic time-series forecasting. It ranks second on GIFT-EVAL among open-source models, only after Chronos-2. It would be useful to support it in Darts like other foundation models (Chronos2 #2943 , TimesFM2.5 #2976).

Apologies for not having the bandwidths for this at the moment. If the maintainer approve this, I recommend this to be a good starter issue for people to get familiar with foundation models.

Describe proposed solution
Add PatchTSTFMModel using Darts FoundationModel + HuggingFaceConnector pattern, with:

• zero-shot inference support,
• probabilistic forecasts via QuantileRegression,
• deterministic median fallback when no likelihood is set,
• fine-tuning support if possible

Describe potential alternatives

Additional context
There are two flavours of PatchTST-FM weights:

• ibm-granite/granite-timeseries-patchtst-fm-r1 is Apache-2.0 and should be the better default for Darts.
• ibm-research/patchtst-fm-r1 is cc-by-nc-sa-4.0 (non-commercial). It can be supported by setting hub_model_name.

The codebase itself ibm-granite/granite-tsfm is Apache-2.0 and compatible with Darts license. We need to provide proper attributions to original license and disclose major changes.

References

• IBM’s newest time-series models cover a full range of enterprise prediction tasks
• arXiv paper
• GitHub ibm-granite/granite-tsfm
• Getting Started Notebook
• HF ibm-research/patchtst-fm-r1
• HF ibm-granite/granite-timeseries-patchtst-fm-r1

cc: @wgifford

[dennisbader]

Hi @daidahao and thanks for this. Yes, sounds good, we can add the model :)