moment

Why we included this project

Teams working with sensor logs, financial series, or operational telemetry often find that a model trained for one task does not transfer well to the next. MOMENT is a pre-trained transformer for general-purpose time-series analysis, covering forecasting, classification, anomaly detection, imputation, and representation learning from a single checkpoint, so you can move between tasks without retraining from scratch. It installs as a Python package, with small, base, and large weights on Hugging Face and notebooks that walk through each supported task. The project also releases the Time-series Pile dataset and separate research code for pre-training your own foundation model on your own data, which matters if you want to go beyond the published weights. For teams that would rather keep one open baseline than maintain a separate model per problem, it is a solid place to start.