Google released TimesFM-3, a 330-million-parameter foundation model for multivariate time-series forecasting, trained on over a trillion real-world and synthetic data points1. The model is available on Hugging Face under a non-commercial license that bars commercial or production use of the default pretrained weights.
TimesFM-3 is Google's first forecasting model natively pre-trained to handle multiple time series with zero-shot generalization. Google research scientists Ayush Jain and Rajat Sen said in the announcement that most real-world forecasting problems are inherently multivariate, where multiple time series and auxiliary external features jointly affect future forecasts. Past sales alone, they explained, tell only part of the story; a useful forecast should also draw on sales of related products, historical foot traffic, and known future events such as weather forecasts, promotions, and holidays.
The model is a decoder-only transformer that segments each time series into patches of 32 data points. It uses two alternating attention layers: one causal within a series and one across all series at a given time step. For inference, TimesFM-3 appends masked placeholder tokens for the entire forecast horizon and fills them in with a single forward pass.
Google shared benchmarks showing TimesFM-3 outperforms all models it compared against, often by a wide margin. Its predecessor, TimesFM-2.5, was state-of-the-art when it launched in September 2025. Google said it will soon replace TimesFM-2.5 as the model powering BigQuery's AI.FORECAST command.
The competitive field has grown denser. Amazon's Chronos-2, Salesforce's Moirai 2.0, and Datadog's Toto 2.0 all shipped over the past year as multivariate-capable forecasting models.
ANALYSIS The non-commercial license creates a gap between the model's benchmark performance and its practical utility for enterprises. Organizations evaluating TimesFM-3 for production workloads would need to wait for a commercial pathway or rely on the BigQuery integration once the swap from TimesFM-2.5 takes effect. The single-forward-pass inference design and 330-million-parameter footprint position TimesFM-3 as a relatively lightweight model compared to general-purpose large language models, a factor that could matter for latency-sensitive forecasting pipelines if commercial access materializes.