TimesFM
Pretrained foundation model for zero-shot time-series forecasting
What it is
TimesFM is a decoder-only pretrained model from Google Research that produces forecasts for time series without per-task training. Version 3.0 handles univariate and multivariate series with past and future covariates and returns quantile forecasts, while 2.5 supports contexts up to 16k steps. Data scientists and analysts load the checkpoints through a Python package with PyTorch, Flax, or MLX backends to forecast demand, traffic, or other sequential metrics, and can fine-tune with LoRA. Note that 3.0 weights carry a non-commercial license, while earlier weights are Apache-2.0.