We tackle the challenge of few-shot time series generation by proposing a unified pretrained model that outperforms state-of-the-art baselines across diverse domains.