SYLIN

TimesFM

by Google · Model · Language & code models

Open-source zero-shot multivariate time-series foundation model, 330M parameters, pretrained on over a trillion time points

AccessOpen weightsNeeds~1.3 GBDownload1.3 GBSize0.33B parametersLicenseCode Apache 2.0; TimesFM 3.0 weights under a non-commercial license

Evidence

  1. Sep 6
    Ranked first among time series foundation models on benchmarks including GIFT-Eval
  2. Sep 6
    Only 1.32 GB on Hugging Face; should fit most consumer devices; GitHub has local run instructions
  3. Sep 5
    Google Research's 330M-parameter decoder-only time-series foundation model, pretrained on over a trillion real and synthetic time points, forecasts zero-shot
  4. Sep 5
    Supports covariates: known future events like promotions, holidays or weather can be fed in to adjust the forecast
  5. Sep 5
    On a synthetic Sydney Pride app-usage test, adding an event flag cut error from about 3,000 to about 1,000 users/day

Cost to run