PatchTST (ETTh1)

This checkpoint is IBM's hosted PatchTST reference model for ETTh1 rather than a broad multi-domain TSFM family card. PatchTST itself is an influential patch-based forecasting architecture, but the weights served here are a dataset-specific reference fit, not a foundation model — so this entry is best understood as a baseline rather than a general-purpose forecaster.

At 616K parameters it uses patch tokenization and Transformer blocks to model long-horizon time-series behavior efficiently. The catch is the training scope: per the official model card, the checkpoint is trained on the ETTh1 train split covering all seven ETTh1 channels, and that same seven-series target shape is what it expects at inference time. It is tuned around the exact ETTh1 setup instead of generic univariate forecasting, so it does not transfer to arbitrary series the way the foundation models here do.

On TSFM.ai reach for it as a narrow, well-understood long-horizon baseline — a fixed point of comparison against the larger foundation models in the catalog — rather than for production forecasting on your own data. When you want the PatchTST design as an actual zero-shot foundation model, use PatchTST-FM R1 instead, which extends the architecture to broad multi-domain forecasting.