Google Research has introduced TimesFM-3, a 330-million-parameter foundation model designed to enable accurate multivariate time series forecasting in a single forward pass. Building on the success of its predecessors, TimesFM-3 represents a significant step forward by adding native support for complex real-world forecasting scenarios that involve multiple related time series and auxiliary external features—addressing a critical gap in practical forecasting applications across retail, finance, healthcare, and manufacturing. Unlike earlier versions limited to univariate forecasting, TimesFM-3 handles multiple coevolving time series simultaneously, capturing cross-series dependencies without requiring task-specific fine-tuning. The model employs an alternating attention architecture combining causal temporal attention (tokens look backward in time within their series) and full variate attention (tokens examine other time series), allowing it to learn complex relationships between data streams. Pre-trained on over 1 trillion time points from both real-world and synthetic data, TimesFM-3 maintains the zero-shot generalization capability that made earlier versions valuable across diverse domains. A key architectural innovation is the non-autoregressive decoding strategy, which generates an entire forecasting horizon in a single pass rather than iteratively predicting one step at a time. This approach reduces latency and error accumulation while handling past covariates (historical features only), past-future covariates (known future signals like scheduled promotions or weather forecasts), and multiple target series. The model produces both point and quantile forecasts, providing probabilistic uncertainty estimates that enable risk-aware decision-making in high-stakes applications.