The 330M-parameter model uses patched time-series inputs and alternating temporal and cross-series attention. Historical signals remain causally constrained, while known future covariates use a lookahead representation. A non-autoregressive decoding strategy produces forecasts in a single forward pass rather than generating each future patch sequentially.
TimesFM-3 is useful for research on demand, sensors, operations, and other multivariate forecasting tasks. It supplies point and quantile predictions and is pretrained on more than one trillion time points. The public code and pretrained weights have different licensing terms; the 3.0 weights are restricted to noncommercial use.
