Machine learning assisted modeling of interatomic potential energy functions is radically changing the field of molecular simulation. With the accumulation of high-quality electronic structure data, it is expected that a model can be pre-trained on all available data and fine-tuned on downstream tasks, enabling rapid deployment and application with minimal additional effort. We propose the Large Atomic Model (LAM) architecture, training methods, fine-tuning strategies, and model knowledge distillation techniques, and preliminarily establish the corresponding workflow. At the model architecture level, we developed the DPA-2 model architecture based on message passing and attention mechanisms, which achieved better results in potential energy function modeling tasks than existing model architectures. We developed multi-task training methods to solve the problem of inability to jointly train first-principles data from different fields, greatly improving the model’s generalization capabilities. Our tests in practical application systems show that even with only the lowest 0.25% of training data, the fine-tuning-distillation model application effect can be fully comparable to the model trained from scratch on the full dataset. The construction of the Large Atomic Model requires an open and collaborative community environment. Therefore, we propose the Open Large Atomic Model Project, OpenLAM, and we welcome the audience to participate in the development and contribution of OpenLAM in any way.