timesfm
Installation
SKILL.md
TimesFM
Overview
TimesFM is a pretrained decoder-only time-series foundation model from Google Research. You pass raw history, it returns a point forecast and quantiles with no fitting. The open-source package timesfm (latest on PyPI: 3.0.2, September 2026) ships two generations of weights, and they have different licenses:
| Model | Python API | Weights license |
|---|---|---|
TimesFM 2.5 (200M parameters, context up to 16k, quantile head up to 1k horizon, no freq argument) |
timesfm.TimesFM_2p5_200M_torch |
Apache-2.0 |
| TimesFM 3.0 (multivariate series, past-only and past-and-future covariates) | timesfm3.TimesFM3Forecaster |
non-commercial license: no commercial or production use of downloaded weights (production use goes through BigQuery ML) |
Pick 2.5 unless the user needs 3.0 features and the use is non-commercial. The 1.x/2.0 API (timesfm.TimesFm, TimesFmHparams, freq=[...]) no longer exists in 3.x; it needs pip install timesfm==1.3.0.