Google Research released TimesFM-3, a foundation model for time series forecasting, on August 31, 2026. Earlier generations projected the future from a single series alone. This one accepts multiple series and external factors at the same time, while keeping the zero-shot convenience of running without additional training. The design takes on the feature engineering work that real-world forecasting used to demand.
One Line Was Never Enough
Think about forecasting sales. If you want to know how much ice cream a chain will sell next month, looking only at past ice cream sales gets you halfway. Sales of related products such as cones and syrups, historical foot traffic, planned promotions, and the weather forecast all shape the actual outcome.
Until TimesFM-2.5, released in September 2025, the models handled only univariate forecasting, meaning the history of one series. Anyone who wanted to use related signals had to build features by hand or bring in a separate machine learning model. The main appeal of a foundation model, that it works immediately without training, faded the moment it reached a production setting.
TimesFM-3 is pre-trained with the multivariate case assumed from the start. It has 330 million parameters and was trained on a time series corpus of more than 1 trillion time points drawn from real and synthetic data. The model is not large by current standards, but the range of information it can absorb is a step wider.
Three Kinds of Input
TimesFM-3 divides the information it accepts into three categories.
The first is multiple targets. Related series can be forecast together, such as several brands of ice cream at once, and both point and quantile forecasts are supported for every target.
The second is past covariates: signals like foot traffic, where the historical values are known but the future values are not.
The third is past-future covariates, and this is where the practical value shows. Promotion dates and weather forecasts are already fixed as future plans. Because the model can be told what will happen during the forecast horizon, it learns from history how promotions move sales and applies that relationship to the scheduled days ahead.
In the example Google published, a univariate model simply extended the weekly pattern forward, while TimesFM-3 with a promotion schedule attached returned a forecast that built in roughly a 20 percent lift on each promotion day. Summed across a month, the gap in projected revenue is not something you can ignore.
The Whole Horizon in a Single Pass
The internal structure follows the decoder-only transformer used in earlier versions. Contiguous runs of 32 time steps are grouped into patches, and per-series normalization absorbs differences in scale. For past-future covariates, each token concatenates the current patch with future patches so the model can look ahead at known signals.
Once tokenized, processing runs as a two-dimensional grid. Attention across time is strictly causal, so a token sees only earlier tokens within its own series. Attention across series is unrestricted, so at any time step a token can look at every other series. Alternating these two layer by layer lets the model learn temporal patterns and cross-series correlations at the same time.
Generation changed as well. Earlier versions produced one patch at a time, which accumulated latency, compounding error, and compute cost. TimesFM-3 uses Contiguous Patch Masking, lining up masked tokens for the full horizon from the outset and filling them all in one forward pass. With no iterative loop, it is simply faster.
The output is not limited to point estimates. For every target at every horizon step, the model returns 9 quantiles from the 10th to the 90th percentile, so forecast uncertainty arrives as a numeric range. In inventory and staffing plans, the spread often matters more than the median, and this output format fits that kind of work.
Top Rank on Three Benchmarks
The evaluation used three public benchmarks: GIFT-Eval, FEV-Bench, and TIME. Among pre-trained foundation models, TimesFM-3 ranked first on both point and probabilistic forecasting metrics. The comparison set included multivariate-capable models such as Chronos-2 and the Toto 2.0 family, along with the previous TimesFM-2.5.
What stands out is that TimesFM-3 matches or beats the other models even in univariate mode, with no covariates and no cross-series information. Support for the multivariate case did not cost anything in the traditional use case. Switching to full multivariate mode then widens the margin further.
Where to Try It Now
TimesFM-3 is available on GitHub and Hugging Face, and BigQuery integration is expected within the coming weeks. Anyone who wants a feel for it sooner can already call TimesFM-2.5 for univariate tasks through the AI.FORECAST command in BigQuery. It runs as an extension of SQL and needs no machine learning expertise, which makes it a reasonable way to test the waters on internal data.
Google notes that TimesFM-3 was developed against adoption in retail, finance, observability, manufacturing, healthcare, and the natural sciences. Time series forecasting gets harder as the number of series grows, and clearing that hurdle without additional training carries real weight for teams that cannot staff a dedicated data scientist.
Summary
TimesFM-3 comes down to three things: pre-training that assumes the multivariate case, generation of the full horizon in a single forward pass, and top rank across three public benchmarks. The ability to feed in information already fixed for the future, such as promotions and weather, is what pays off in practical forecasting. Since the model is already on GitHub and Hugging Face, the fastest way to judge it is to run your own data through univariate and multivariate modes and see how much preparing covariates is actually worth.
