Abstract. Climate models are hindered by the need to conceptualize and then parameterize complex physical processes that are not explicitly numerically resolved and for which no rigorous theory exists. Machine learning and artificial intelligence methods (ML/AI) offer a promising paradigm that can augment or replace the traditional parametrized approach with models trained on empirical process data. We offer a flexible and efficient framework, TorchClim, for inserting ML/AI physics surrogates that respect the parallelization of the climate model. A reference implementation of this approach is presented for the Community Earth System Model (CESM), where the authors substitute moist physics and radiative parametrization of the Community Atmospheric Model (CAM) with an ML/AI model. We show that a deep neural network surrogate trained on data from CAM itself can produce a stable model that reproduces the climate and variability of the original model, albeit with some biases. This framework is offered to the research community as an open-source project. The new framework seamlessly integrates into CAM's workflow and code-base and runs with negligible added computational cost, allowing rapid testing of various ML physics surrogates. The efficiency and flexibility of this framework open up new possibilities for using physics surrogates trained on offline data to improve climate model performance and better understand model physical processes.