The paper introduces a novel framework called Co-Boosting to enhance one-shot federated learning (OFL). OFL aims to train a global server model using only a single communication round, making it more practical and efficient compared to multi-round communication methods. The performance of the server model in OFL is closely tied to the quality of synthesized data and the ensemble model. To address this, Co-Boosting leverages the current ensemble model to generate higher-quality samples adversarially, which are then used to improve the ensemble model by adjusting the weights of each client model. This mutual enhancement process is repeated periodically, leading to high-quality data and ensemble models. Extensive experiments on various datasets demonstrate that Co-Boosting outperforms existing baselines, achieving significant improvements in accuracy, especially under non-IID settings and with heterogeneous client models. The method is also practical for contemporary model-market scenarios, as it does not require adjustments to local training, additional data or model transmissions, and accommodates diverse client model architectures.The paper introduces a novel framework called Co-Boosting to enhance one-shot federated learning (OFL). OFL aims to train a global server model using only a single communication round, making it more practical and efficient compared to multi-round communication methods. The performance of the server model in OFL is closely tied to the quality of synthesized data and the ensemble model. To address this, Co-Boosting leverages the current ensemble model to generate higher-quality samples adversarially, which are then used to improve the ensemble model by adjusting the weights of each client model. This mutual enhancement process is repeated periodically, leading to high-quality data and ensemble models. Extensive experiments on various datasets demonstrate that Co-Boosting outperforms existing baselines, achieving significant improvements in accuracy, especially under non-IID settings and with heterogeneous client models. The method is also practical for contemporary model-market scenarios, as it does not require adjustments to local training, additional data or model transmissions, and accommodates diverse client model architectures.