Google Research released TimesFM-3, a 330-million-parameter time series forecasting model trained on over 1 trillion data points. The big shift from earlier versions: it handles multivariate forecasting—meaning it can work with multiple related time series and external information at once, not just a single data stream. Real forecasting problems need this. Predicting ice cream sales requires knowing not just past sales but related product sales, foot traffic, weather forecasts, and promotion schedules. TimesFM-3 processes time series by chunking them into 32-step patches, then uses a two-layer attention system: causal temporal attention (tokens look backward in time within their own series) and full variate attention (tokens see all other series at each time step). The clever part is how it handles future information—it uses a "lookahead" strategy for known future events like holidays or promotions, letting those signals influence predictions while keeping unknown future values masked. The model generates entire forecasts in one pass rather than iteratively, which cuts latency and error accumulation. It predicts nine quantile levels (10th through 90th percentile) for each target series, giving a full uncertainty picture instead of just point estimates. In the ice cream example, when you feed in a promotion schedule as a "past-future covariate," the model learns the historical relationship between promotions and sales, then applies that pattern to upcoming promotion days—catching a ~20% sales bump that a standard univariate model would miss completely. On three major benchmarks (Gift-Eval, FEV-Bench, and Time), TimesFM-3 ranked first among foundation models for both point forecasting accuracy and probabilistic metrics. Even running in univariate mode—ignoring multivariate information—it matched or beat competing models. When multivariate features were included, the gap widened further. The model is available on GitHub and Hugging Face now, with BigQuery integration coming soon.