Google Releases TimesFM-3, a New Time-Series Forecasting Model, Under Non-Commercial License

Google Introduces TimesFM-3

Google has launched TimesFM-3, a new time-series forecasting model. This model features 330 million parameters and was trained using over a trillion real-world and synthetic data time points. It is now accessible on Hugging Face, though it is released under a non-commercial license.

Advancements in Forecasting

TimesFM-3 represents a new generation of forecasting models capable of ingesting multiple time series. Unlike previous models, it was natively pre-trained to handle multivariate data and offers zero-shot generalization. This allows it to forecast multiple related time series in parallel and incorporate diverse historical data, such as past foot traffic or sales of related products, alongside external factors like weather or promotions.

Performance Benchmarks

In internal benchmarks shared by Google, TimesFM-3 demonstrated superior performance compared to other leading models in the field, including Salesforce’s Gift-Eval, Amazon/AutoGluon’s FEV-Bench, and Time. The rapid pace of development in this area is highlighted by the fact that TimesFM-2.5, which was state-of-the-art in September 2025, now ranks at the bottom of these benchmarks.

Architectural Details

TimesFM-3 utilizes a decoder-only transformer architecture, similar to its predecessors. It processes time series by dividing them into patches of 32 data points, treating these patches like tokens in a language model. The model's architecture includes two alternating types of attention layers: one that looks backward causally within a single series and another that examines all series simultaneously.

Implications for the Industry

While the non-commercial license restricts immediate business applications, the release of TimesFM-3 signifies a notable advancement in time-series forecasting. Its ability to handle complex, multivariate data with zero-shot generalization sets a new benchmark for accuracy and efficiency in predictive analytics, pushing the capabilities of AI in business and research contexts.