Google released TimesFM-3, a 330M-parameter time series foundation model with native multivariate forecasting (non-commercial license)

TimesFM-3 is the third generation of Google Research's zero-shot forecasting model, and the main change from 2.5 is that it handles multivariate inputs natively instead of being limited to a single series' own history. It supports multiple simultaneous targets, past-only covariates, and past-future covariates (things like holidays or planned promotions where future values are known), all without fine-tuning.

Architecturally it's a decoder-only transformer with 20 layers at model dim 1280 and 16 heads, patching 32 contiguous time steps per token, and alternating two attention types per layer: causal attention across time within a series, and full attention across series at a given time step. Forecasts are generated in one forward pass rather than autoregressively — the model appends masked placeholder tokens for the whole horizon and fills them in simultaneously, with past-future covariates left unmasked so their known values stay visible. It outputs 9 quantiles (10th–90th percentile) per target per horizon step.

Pretraining used GiftEvalPretrain (minus fev-bench overlaps), Wikipedia pageviews through Nov 2023, Google Trends queries through end of 2022, plus synthetic data, totaling over 1 trillion time points. Google reports best average rank on Gift-Eval, FEV-Bench, and Time against Chronos-2, Toto 2.0, and TimesFM-2.5, and claims the univariate-only mode already matches or beats those baselines before covariates are added.

Worth flagging: the weights are under the TimesFM Non-Commercial License v1.0, so this isn't a drop-in for production use the way some other releases are. PyTorch weights are on Hugging Face and GitHub now; BigQuery integration is listed as coming later.

- Research Blog: https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting/