Time-MoE introduces a breakthrough architecture for time series forecasting that scales to billion-parameter models while remaining computationally efficient.

• Combines transformer architecture with mixture-of-experts to handle massive time series data at scale
• Enables unified pre-training across diverse time series domains
• Achieves superior forecasting capabilities at lower computational costs than previous approaches
• Designed for real-world engineering applications where accurate prediction is critical

This research represents a significant advancement for engineering applications requiring time series forecasting, from predictive maintenance to demand planning, by making large-scale foundation models practical for industrial deployment.

Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts