AI Briefingtimesfm-3.0-pytorch: Google's Foundation Model for Time Series Forecasting with 0.3B Parameters
google/timesfm-3.0-pytorch
About the project
A foundation model from Google Research that lowers the barrier to entry for time series data analysis. It focuses on forecasting future values by applying pre-trained weights without complex modeling.
It combines the Stacked Mixing Transformer architecture with Variate Attention and CPM Iterative RevIN to enhance precision. With 20 layers and a model dimension of 1280, it maintains a scale of 0.3B parameters while supporting efficient inference.
Pre-trained on diverse sources such as Wikipedia page views, Google Trends, and synthetic data, it is robust against common time series patterns. It provides default settings of 32 context patches and 64 forecast horizon patch lengths.
It includes official PyTorch-based weights and configurations for easy integration into existing pipelines. However, due to the application of a non-profit license, conditions should be verified before commercial use.
google/timesfm-3.0-pytorch
The original page has no description.
time-series-forecasting
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.