AI BriefingTimesFM (GitHub Repository)
Key point
Google Research has released **TimesFM 2.5**, a foundation model for time series forecasting.
Details
TimesFM (Time Series Foundation Model) is a decoder-only foundation model for time series forecasting developed by Google Research.
The latest version, TimesFM 2.5, brings major improvements in model efficiency and performance compared to the previous version, 2.0. The key changes are as follows:
- Parameter optimization: Reduced the number of parameters from 500M to 200M to improve efficiency.
- Context extension: Expanded the supported context length from the previous 2,048 up to 16k.
- Quantile forecasting: Supports continuous quantile forecasting up to a 1k horizon via a 30M quantile head.
- Feature improvements: Removed the
frequencyindicator, added new forecasting flags, and reintroduced covariate support via XReg.
Users can choose either the PyTorch or Flax backend depending on their operating environment, and installation is made simple via the uv package manager.
This summary 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 summary errors, attribution issues, or removal requests via Contact.
Related
Google Releases TimesFM-3 for Multivariate Time Series Forecasting