Google Research · 2026-08-31 · major
TimesFM-3 — Google's forecasting model handles many series at once
TimesFM-3 is a 330M-parameter time-series foundation model from Google Research that forecasts several linked series in one forward pass. It ranked first on GIFT-Eval, FEV-Bench and TIME among pre-trained forecasting models.
Google Research's 330M forecasting model predicts many linked time series in a single pass, with no fine-tuning.
Key specs
| Parameters | 330M |
|---|---|
| Pretraining corpus | 1T+ time points |
Quick facts
| Maker | Google Research |
|---|---|
| What's new | Native multivariate forecasting in one forward pass |
| Covariates | Past-only and past-plus-future |
| Outputs | Point forecast plus 9 quantiles (10th–90th) |
| Benchmarks | 1st on GIFT-Eval, FEV-Bench and TIME |
| License | Apache-2.0 code, non-commercial weights |
| Availability | GitHub and Hugging Face; BigQuery coming |
What is it?
TimesFM-3 adds native multivariate forecasting: it predicts several series that move together in one shot, rather than running a separate model per series. Google Research trained the 330M-parameter model on more than 1 trillion time points drawn from real-world and synthetic data. The weights are on Hugging Face and the code is on GitHub.
How does it work?
A decoder-only transformer splits each series into 32-step patches and normalizes them per series. Attention then alternates between two modes: causal temporal attention, where a token only sees the past of its own series, and full variate attention, where a token sees every series at the same timestep. A method called Contiguous Patch Masking lets the model emit a complete forecast in one forward pass instead of stepping through it.
Why does it matter?
Demand planning, capacity forecasting and sensor monitoring almost never involve a single lonely series. TimesFM-3 takes past covariates and known future covariates, and returns both a point forecast and nine quantiles from the 10th to the 90th percentile, so teams get an uncertainty range and not just a line. That is a strong zero-shot baseline without training anything per dataset.
Who is it for?
data scientists and forecasting teams
Frequently asked questions
- Can I use TimesFM-3 in a commercial product?
- Not with the default weights. TimesFM-3's pretrained checkpoint on Hugging Face ships under the TimesFM Non-Commercial License v1.0, which limits it to non-commercial, non-production use. The google-research/timesfm code repository itself is Apache-2.0, so the library is free to use, but the released weights are the part that carries the restriction.
- How does TimesFM-3 differ from earlier TimesFM versions?
- TimesFM-3 adds native multivariate forecasting and covariate support, which earlier TimesFM releases did not have. Google Research says the 330M-parameter model jointly predicts multiple coevolving series and captures dependencies between them, while keeping the zero-shot generalization of its predecessors. Even in univariate mode, without covariates, TimesFM-3 still outperformed the competing foundation models.
- Does TimesFM-3 need fine-tuning on my own data?
- No. TimesFM-3 is a zero-shot model: Google Research reports it topped GIFT-Eval, FEV-Bench and TIME in both point and probabilistic forecasting without task-specific fine-tuning. You supply the history, optionally past covariates and known future covariates such as planned promotions, and TimesFM-3 returns the forecast directly.
- Where can I get TimesFM-3 today?
- TimesFM-3 is published on GitHub at google-research/timesfm and on Hugging Face as google/timesfm-3.0-pytorch. Google Research says a BigQuery integration is landing in the coming weeks, so teams that keep their series in a warehouse will be able to forecast in place rather than exporting data to a separate serving stack.
Try it
Model id: google/timesfm-3.0-pytorch