A foundation model for text is judged on prose nobody wrote for it. A foundation model for time series makes the same bet on a signal it was never trained on, which is the whole reason the word pretrained sits in the description.
Classical forecasting fits a model to your series first. A pretrained one arrives already fitted to other series, so the question is how far that transfers to yours.
Naming the lab tells you the training corpus and the evaluation live in a paper somewhere, which is more than most repos on this board can say.
No horizon, no accuracy, no dataset, no comparison against a baseline. The trending description names the model and its job and stops there.
Concept mock, not affiliated with the project. The GitHub Trending description on 4 September 2026 reads "TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting." The chart is an illustrative drawing: the solid line, the shaded band, and the dashed centre line were drawn by hand to show the shape a forecaster outputs. Both axes are unlabelled and no value, horizon, interval, or accuracy figure is represented, because that sentence publishes none.