TimesFM 3.0 can be used in Google Cloud BigQuery via AI.FORECAST for multivariate tasks, or run from the open checkpoint google/timesfm-3.0-pytorch on Hugging Face for research or non-commercial work, with both paths delivering probabilistic outputs and covariate influence.

Usage options include BigQuery with AI.FORECAST for multivariate forecasting and the open checkpoint for experimentation, enabling probabilistic forecasts and covariate effects.

Key features include zero-shot forecasting, native multivariate support, probabilistic forecasts across nine quantiles, fast one-pass decoding, and a compact model suitable for BigQuery and other platforms.

The model has 330 million parameters, was pretrained on over a trillion time points (real and synthetic), and can forecast multiple related time series simultaneously while incorporating covariates such as promotions, weather, and holidays.

Looking ahead, TimesFM 3.0 signals a shift toward a single general model capable of handling multiple forecasting tasks, easing access to AI-powered analytics across industries via cloud platforms and certifications.

TimesFM 3.0 uses a decoder-only transformer architecture with four processing steps: patching and normalizing data, building tokens with covariates, alternating causal temporal and full variate attention, and forecasting the entire horizon in one pass with Contiguous Patch Masking.

Limitations include that forecasts are not guarantees, covariates must be reliable, zero-shot results may improve with domain knowledge and data cleaning, and the BigQuery feature is in preview with evolving licensing; validate on your data.

TimesFM 3.0 is Google Research’s third-generation Time Series Foundation Model, released in August 2026, designed for zero-shot multivariate forecasting without retraining on user data.

Real-world use cases span retail demand and promotions, transportation data like NYC Citi Bike, energy, finance, monitoring, and marketing analysis, with potential to pair forecasts with certifications for broader skills.

Compared with TimesFM 2.5, TimesFM 3.0 adds multivariate forecasting, supports past and future covariates, and decodes the entire horizon in one pass, outperforming on benchmarks GIFT-Eval, FEV-Bench, and TIME in both point and probabilistic forecasts.

Licensing notes state pretrained weights are non-commercial for non-commercial, non-production use; commercial/production use is allowed via Google Cloud services like BigQuery ML.