Google drops TimesFM 3.0 with native multivariate forecasting

TimesFM 3.0 is Google Research's open-source time-series foundation model designed to forecast real-world numerical trends like retail demand, logistics volume, and energy load right out of the box.

While previous releases like TimesFM 2.5 were strictly univariate and could only inspect a single historical curve at a time, TimesFM 3.0 natively handles multivariate forecasting. It can analyze multiple correlated signals alongside scheduled future events, resolving the entire forecast horizon in a single forward pass without requiring custom fine-tuning.

What it does

In practice, native multivariate support means the model looks at real-world forecasting like an actual operator instead of reading a isolated line chart.

If you want to forecast next week's ice cream sales, for instance, you no longer rely solely on past ice cream sales. TimesFM 3.0 can simultaneously ingest complementary product curves like cones and syrup, check foot-traffic data, and factor in upcoming variables like planned promotional calendars, holiday shifts, and local weather forecasts.

Google tested the architecture across standard benchmarks including GIFT-Eval, FEV-Bench, and TIME. Across both point and probabilistic forecasting metrics, TimesFM 3.0 claimed the top spot among foundation models, outperforming alternatives like Chronos-2 and Toto 2.0.

Specs and how to run it

The model is remarkably compact. TimesFM 3.0 packs 330 million parameters and comes out to just 1.32GB on disk.

Google pre-trained it on a massive collection of more than 1 trillion time points. Because of its small footprint, it runs smoothly on as little as 1GB of VRAM, making it fully deployable on consumer laptops and edge workstations for zero-shot inference without retraining.

Google uploaded official PyTorch weights and configuration files to Hugging Face under the repository google/timesfm-3.0-pytorch, with setup instructions provided on GitHub. The critical catch is its licensing: Google released TimesFM 3.0 under the TimesFM Non-Commercial License v1.0, strictly limiting usage to research and non-commercial evaluation.

Why it matters

If you work with operational time series, financial curves, or supply chain data, TimesFM 3.0 is worth spinning up immediately. Delivering competitive zero-shot predictions on 1GB of VRAM with single-pass speed eliminates the overhead of constantly retraining brittle bespoke statistical setups.

The only real roadblock is the non-commercial license, which keeps production business pipelines on the sidelines for now. Even so, it proves that foundation models do not need billions of parameters to dominate a niche: a lean 330M-parameter architecture backed by high-quality pre-training data can deliver frontier accuracy on everyday consumer hardware.