Google released TimesFM-3 on September 12, 2026, saying it predicts future events from time series data like daily sales figures. It is the company's first forecasting update since the launch of TimesFM-2.5 in September 2025.
Google reported TimesFM-3 has 330 million parameters and was trained on over one trillion data points, measured on real and synthetic time series. That compares with its predecessor, TimesFM-2.5, which had fewer parameters and was trained on a smaller dataset.
TimesFM-3 is built on a Transformer architecture and targets use cases like retail sales forecasting, incorporating related data such as weather and discount schedules. Availability begins on GitHub and Hugging Face, initially for developers and data scientists.
"Real-world forecasts rarely depend on a single variable," said Jonathan Kemper, Google Research. "TimesFM-3 improves predictions by integrating multiple data sources like weather and discount campaigns."
The announcement follows Google DeepMind's release of WeatherNext Cyclones, an open-source AI system for tropical cyclone prediction. Google said the release reflects its ongoing commitment to advancing forecasting models across industries.
Google did not say when TimesFM-3 will be integrated into BigQuery, and it raised the open question of how the model will handle real-time data updates. The company plans to add it to BigQuery in the coming weeks.
Source: thedecoder
