Show HN: T0: We trained a 256M SOTA open-weights time-series forecasting model (github.com)

The Forecasting Company has unveiled T0, a 256M state-of-the-art open-weights time-series forecasting model designed to deliver probabilistic forecasts using historical data and known future covariates. This foundation model allows users to enhance forecasting accuracy by leveraging diverse inputs, including calendar features and external factors, making it particularly useful for applications like predicting national electricity demand and drug reimbursement trends in France. The model, built on a patch transformer architecture, supports forecasting multiple horizons simultaneously, with capabilities for univariate and multivariate analyses. Significantly, T0 introduces features such as advanced quantile predictions and the ability to efficiently handle batch inputs through its native TimeSeries integration. Technical highlights include the implementation of rotary embeddings and a flexible configuration system that adapts to different data contexts. The model is also available for local inference through PyTorch and Apple’s MLX runtime, alongside ONNX compatibility for broader deployment options. T0 not only democratizes access to sophisticated forecasting tools but also sets a precedent for open-source collaboration in the AI/ML community by integrating concepts from existing models, encouraging further innovation in time-series analysis.