Yug
An open time series foundation model from Birla AI Labs for zero-shot probabilistic forecasting across domains and frequencies.
Yug is a 272M-parameter, patch-based decoder that forecasts unseen series without any task-specific training. It produces quantile forecasts over a context of up to 2,048 points, handles optional covariates, and scores a CRPS of 0.529 and a MASE of 0.744 on GIFT-Eval. You can install it with pip install yug.
I worked on the synthetic data used to pretrain the model. This builds on my research into synthetic corpus composition for time series pretraining, which was presented at two ICML 2026 workshops.
Released by Birla AI Labs with Aaditya Jain, Debdeep Sanyal, Dhruv Kumar, Murari Mandal and Saurabh Deshpande.