Abstract: With respect to the problem due to the different multi-GPU types in cone-beam CT reconstruction, a model was proposed based on heterogeneous multi-GPUs. Using FDK algorithm for reconstruction, the model allocated tasks according to the computing capacity of each GPU, ensuring the balance among GPUs. Massive data image reconstruction was achieved by data flow decomposition. The implementation of the method was carried out based on CUDA, including multi-GPUs parallel computing using data flow management and asynchronous function and the design of the kernel function in filtering and back-projection. The model was tested on the high precision industrial CT system. The results illustrate that the reconstruction model is accurate and effective, taking full advantage of heterogeneous multi-GPUs, and is considerably effective compared to conventional methods.