TimesFM-3 Natively Supports Multivariate Zero-Shot Forecasting, but Its Weights Restrict Commercial and Production Use

Google Research released TimesFM-3 on August 31, expanding its previously univariate-focused general-purpose time-series model into a native multivariate forecaster. The model has 330 million parameters and was pretrained on more than one trillion real, synthetic, and augmented time points. Its inputs can simultaneously include multiple forecast targets, historical-only covariates, and known-future signals such as weather forecasts and promotion schedules, without requiring fine-tuning for each dataset.

The model first divides each series into patches of 32 time steps and normalizes each series independently. Inside the Transformer, the data is arranged as a two-dimensional “time × variable” grid: causal temporal attention can access only the past of the same series, while full variable attention exchanges information among different series at the same time position. The two are stacked alternately to separate temporal dependencies from cross-series relationships. For known-future covariates, tokens are concatenated with subsequent patches, allowing the model to use scheduling information without leaking target values.

Another change is non-autoregressive decoding. TimesFM-3 uses Contiguous Patch Masking to append masked positions after the observed data and generates the entire forecast horizon at once instead of advancing one patch at a time. In theory, this can reduce latency and the accumulation of errors. For every target and time step, the model outputs nine quantiles from the 10th through the 90th percentile, expressing uncertainty in addition to point estimates.

Google says the model ranks first across the GIFT-Eval, fev-bench, and TIME benchmark suites, and that its multivariate mode outperforms its own univariate configuration. However, the figures are still reported primarily by the development team, and the pretraining data also includes public time series. Practitioners should therefore run their own backtests using application-specific temporal splits while accounting for distribution drift and covariate availability. More importantly, although the repository’s code is licensed under Apache 2.0, the TimesFM-3 weights are covered by a separate noncommercial license that explicitly prohibits commercial and production use. Hugging Face also does not yet offer a hosted inference provider for the model. BigQuery integration is expected within weeks; once available, users should verify the licensing terms, service quotas, and whether its results align with those of the local version.