google-research/timesfm +3,203 ★ this week + Follow

What it actually is

TimesFM is a pretrained decoder-only time-series foundation model from Google Research, distributed as a Python package (timesfm[torch] or timesfm[flax]) for zero-shot and fine-tuned forecasting. It ships both PyTorch and Flax/JAX implementations, with core normalisation and utility logic duplicated across src/timesfm/torch/util.py and src/timesfm/flax/util.py (e.g. revin, update_running_stats), plus covariate/regressor support via an XReg library present in both v1/src/timesfm/xreg_lib.py and src/timesfm/utils/xreg_lib.py. It targets practitioners needing a general-purpose forecasting model without training from scratch, and is embedded in Google's own products (BigQuery ML, Connected Sheets, Vertex Model Garden).

Why it's growing

The star bump coincides with the TimesFM-3.0 release (v3.0.0, 2026-08-28) and a cluster of commits actively fixing covariate handling bugs (PR #483, "keep past-future covariate windows aligned when the context is truncated") and README updates announcing the 3.0 model. The repository's integration into Google 1P products (BigQuery ML, Sheets, Vertex Model Garden) cited in the README is a plausible pull for adoption, though the graph facts don't directly quantify external usage — the growth is best explained by the visible release cadence and active bug-fixing rather than any single documented viral event.