Google Research released TimesFM-3, a time-series forecasting model that predicts several related metrics in a single forward pass instead of one series at a time. The model has 330 million parameters and was pretrained on more than 1 trillion time points spanning real and synthetic data, according to Google. It requires no task-specific fine-tuning to run on a new forecasting problem.
That distinction matters more than it sounds. Most forecasting tools in production today handle one series at a time: yesterday’s sales predict tomorrow’s sales, and nothing else enters the calculation. TimesFM-3 is built to ingest a set of related series together, plus two kinds of side information: past-only signals like historical foot traffic, and past-future signals like a promotion calendar or a weather forecast that is already known. Google says the model uses a “lookahead” token construction so it can see those future known events while forecasting.
Forecasting is arguably the most common quantitative job inside an ordinary company: demand planning, staffing schedules, warehouse capacity, cash flow. Those problems have historically needed a data science team to build and retrain a bespoke model for each metric. A 330-million-parameter model that needs no fine-tuning is small enough to run close to where that work actually happens, inside a retailer’s planning tool or a finance team’s spreadsheet pipeline, rather than requiring a dedicated ML engineering effort.
Architecturally, TimesFM-3 keeps the decoder-only transformer design of its predecessors but alternates two attention passes. Causal temporal attention lets a token look backward across time within its own series. Full variate attention lets a token look across every other series at the same time step, which is how the model learns that a promotion in one product line affects sales in another. Earlier versions, by Google’s account, built a forecast in sequential chunks, which the company says cost latency and let errors accumulate; TimesFM-3 instead masks the full future horizon and fills it in one pass, producing nine quantiles per target at every step for a probabilistic forecast rather than a single number.
On the benchmark side, Google evaluated the model against Gift-Eval, FEV-Bench, and the Time leaderboard, and reports that TimesFM-3 ranks first among pretrained foundation models on both point and probabilistic accuracy across all three. The comparison set includes Chronos-2 and the Toto 2.0 family. These are Google’s own reported rankings; the release does not include an independent third-party evaluation, and Google has not published raw score tables alongside the summary rank charts. In Google’s own illustrative example, a retail promotion schedule fed in as a past-future covariate let the model anticipate roughly a 20 percent sales bump on promotion days, a pattern the company says a univariate model misses entirely.
TimesFM-3 is available now on GitHub and Hugging Face. Google says BigQuery integration is coming in the following weeks, which would put multivariate forecasting behind the same AI.FORECAST SQL command that already exposes the univariate TimesFM-2.5 to analysts without a machine learning background.
Teams running separate forecasting models for related metrics (SKU-level demand, regional staffing, multi-branch cash flow) should benchmark TimesFM-3’s zero-shot multivariate mode against their current bespoke pipelines before the next planning cycle; the case for it is strongest wherever a known future event, not just history, should be moving the forecast.
Google Research published the TimesFM-3 announcement on its research blog on August 31, 2026.
