Overview
TimesFM-3 is Google Research’s zero-shot foundation model for univariate and multivariate time-series forecasting. It has 330 million parameters and was pretrained on more than one trillion real-world and synthetic time points. Google announced it on August 31, 2026. (research.google)
Capabilities and uses
- Jointly forecasts multiple related target series, with point forecasts and nine quantiles from the 10th through the 90th percentile.
- Supports historical-only covariates and covariates known for both the historical and forecast periods, such as planned promotions or holidays.
- Generates the forecast horizon in a single forward pass using a decoder-only transformer with alternating temporal and cross-series attention.
- Potential applications include retail demand, finance, manufacturing, healthcare, observability and scientific forecasting. Google reports that the model ranked first among evaluated pretrained foundation models on the GIFT-Eval, FEV-Bench and TIME benchmarks; results for a particular task can differ. (research.google)
- Google provides downloadable PyTorch weights and offers TimesFM 3.0 in BigQuery for forecasting; BigQuery also supports univariate evaluation and anomaly detection. The BigQuery integration is in preview. (huggingface.co)
Limitations and licensing
- Forecasts are uncertain and should be evaluated on representative data. Future covariates must be supplied if they are to inform predictions; genuinely unknown future events cannot be supplied as known inputs.
- In BigQuery, TimesFM 3.0 accepts a context window of 64–2,048 time points in increments of 32; additional input points beyond 2,048 are ignored. This is a BigQuery-specific limit. (docs.cloud.google.com)
- The downloadable weights are subject to the TimesFM Non-Commercial License v1.0, which limits them to non-commercial, non-production use. BigQuery use is governed separately by Google Cloud terms and permits commercial use. (huggingface.co)
Pricing
Checked November 16, 2026. Downloadable weights are royalty-free for permitted non-commercial use; self-hosting can still incur infrastructure costs. (huggingface.co)
For TimesFM 3.0 in BigQuery, preview usage before token-billing enforcement is charged using BigQuery slots for Enterprise or Enterprise Plus, or bytes processed under on-demand pricing. Applicable slot and byte rates depend on the BigQuery pricing configuration; a single current rate is not verified here. There is no separately verified input/output token rate for this preview period. (docs.cloud.google.com)
Google lists a target enforcement date of December 1, 2026 for BigQuery model-token billing: USD $0.05 per million input tokens and USD $0.20 per million output tokens. These are future rates, not the preview rates on the query date. Google defines a time-series data point as one token for this pricing calculation; non-model parts of the query are additionally charged by slots or bytes processed. (cloud.google.com)