Publication date: 1 September 2026
Google Research released TimesFM-3 on 31 August 2026, a 330-million-parameter foundation model for zero-shot time-series forecasting. It is Google’s first TimesFM model trained natively for multivariate forecasting, so it can use related series and known future variables without task-specific fine-tuning.
The model is available through GitHub and Hugging Face now. The most important business caveat is the licence: the TimesFM-3 weights are restricted to non-commercial and non-production use.
What can TimesFM-3 do?
Earlier TimesFM releases focused mainly on univariate forecasting, where one series is predicted from its own history. TimesFM-3 can forecast several related target series together and use additional variables that may improve the prediction.
Google gives a retail example. A sales forecast can use related product demand, historical foot traffic, planned promotions and future weather instead of looking only at past sales. The same pattern applies to finance, manufacturing, observability and healthcare.
TimesFM-3 also supports probabilistic forecasting. Instead of producing only one expected value, it can return quantiles that describe a range of possible outcomes. That is useful when a planning team needs to understand uncertainty rather than work from one point estimate.
TimesFM-3 specifications, access and licence
| Specification | TimesFM-3 |
|---|---|
| Parameters | 330M |
| Training corpus | More than 1 trillion time points |
| Forecasting | Univariate and multivariate |
| Covariates | Past-only and known future variables |
| Official weights | Hugging Face |
| Source code | GitHub |
| Weights licence | TimesFM Non-Commercial License v1.0 |
Google’s Hugging Face checkpoint is about 1.32 GB and uses a 20-layer transformer architecture. The model supports up to 32 variates in the published configuration.
The licence is the main restriction. Google’s repository code remains under Apache 2.0, but the TimesFM-3 pretrained weights use a separate licence that permits non-commercial and non-production use only. A company cannot treat the public checkpoint as a drop-in production forecasting model.
Google says a BigQuery integration is planned for the coming weeks. That may become the more relevant route for business users, depending on the terms and availability of that integration.
How does TimesFM-3 compare with other forecasting models?
Google evaluated TimesFM-3 on GIFT-Eval, FEV-Bench and TIME. The company reports that TimesFM-3 ranked first among pretrained foundation models on both point and probabilistic forecasting metrics across all three benchmarks.
The comparisons include multivariate-capable models such as Chronos-2 and the Toto 2.0 family, plus TimesFM-2.5. Google also reports that TimesFM-3 remains competitive when forced into univariate mode, then improves further when cross-series information and covariates are enabled.
These are Google-reported benchmark results. The release post shows rankings rather than one simple business metric, and the public weights cannot currently be used in commercial production. Teams should therefore test accuracy on their own demand, capacity or inventory data before drawing conclusions from the leaderboard.
What is missing from the TimesFM-3 release?
The biggest missing piece is a production path for the public weights. The model is technically downloadable but legally restricted for commercial deployment, which sharply limits its immediate value for companies evaluating self-hosted forecasting.
Google also does not publish a simple cost comparison against a conventional forecasting pipeline. A 330M model is modest by modern foundation-model standards, but operational value depends on forecast accuracy, inference cost and how much manual feature engineering it replaces.
What this means
Worth testing now for data teams, but not deploying commercially from the public checkpoint. A retail or supply-chain team can use TimesFM-3 to evaluate whether promotions, weather and related products improve forecasts without building a custom multivariate model first.
The first test should compare it with your existing baseline on real planning decisions. Measure forecast error, calibration and downstream inventory or staffing impact. If Google provides a commercially usable BigQuery route, TimesFM-3 could become much more relevant for production teams.
For more information, visit the official announcement of TimesFM-3 on the Google Research website.
