This paper introduces Timer, a large time series model (LTSM) that leverages generative pre-training to achieve state-of-the-art performance in time series forecasting, imputation, and anomaly detection. The model is trained on a large-scale dataset with up to 1 billion time points, and it uses a unified sequence format to handle heterogeneous time series data. Timer is designed as a generative pre-trained Transformer, enabling it to generalize across various tasks and data scarcity scenarios. The model's architecture is inspired by large language models, featuring autoregressive generation and scalability. The paper evaluates Timer on benchmark datasets, demonstrating its effectiveness in forecasting, imputation, and anomaly detection, even with limited data. It also shows that Timer outperforms state-of-the-art small models in these tasks. The model's scalability is validated through experiments that increase both model size and data scale, leading to improved performance. Additionally, Timer exhibits zero-shot forecasting capabilities, allowing it to perform well without fine-tuning on specific tasks. The study highlights the importance of pre-training on large-scale time series data for developing robust and generalizable models. It also addresses challenges in handling heterogeneous time series data and proposes a unified sequence format to facilitate this. The paper contributes to the development of large time series models by providing a comprehensive dataset, a unified sequence format, and a generative pre-trained Transformer architecture. The results show that Timer is a promising approach for time series analysis, offering scalability, versatility, and performance in data-scarce scenarios. The paper also emphasizes the need for further research into the scalability and generalization of large time series models.This paper introduces Timer, a large time series model (LTSM) that leverages generative pre-training to achieve state-of-the-art performance in time series forecasting, imputation, and anomaly detection. The model is trained on a large-scale dataset with up to 1 billion time points, and it uses a unified sequence format to handle heterogeneous time series data. Timer is designed as a generative pre-trained Transformer, enabling it to generalize across various tasks and data scarcity scenarios. The model's architecture is inspired by large language models, featuring autoregressive generation and scalability. The paper evaluates Timer on benchmark datasets, demonstrating its effectiveness in forecasting, imputation, and anomaly detection, even with limited data. It also shows that Timer outperforms state-of-the-art small models in these tasks. The model's scalability is validated through experiments that increase both model size and data scale, leading to improved performance. Additionally, Timer exhibits zero-shot forecasting capabilities, allowing it to perform well without fine-tuning on specific tasks. The study highlights the importance of pre-training on large-scale time series data for developing robust and generalizable models. It also addresses challenges in handling heterogeneous time series data and proposes a unified sequence format to facilitate this. The paper contributes to the development of large time series models by providing a comprehensive dataset, a unified sequence format, and a generative pre-trained Transformer architecture. The results show that Timer is a promising approach for time series analysis, offering scalability, versatility, and performance in data-scarce scenarios. The paper also emphasizes the need for further research into the scalability and generalization of large time series models.