Considering the computing heterogeneity, devices with insufficient computing resources may increase the model training latency of semantic encoder (SE) and semantic decoder (SD) in the task-oriented communication network. In this paper, we propose a cooperative model training (CMT) framework based on federated learning for the training of SE and SD models among computing-heterogeneous devices. First, we design a resource allocation scheme to reduce the computing and communication latency of the proposed framework. Then, we analyze the convergence performance of our CMT framework in the training absence of computing-weak devices. Numerical experi-ments validate that the proposed CMT framework outperforms the benchmarks by obtaining a 6% gain in training accuracy and a 30 % gain in latency. Moreover, the CMT framework achieves nearly indistinguishable performance with centralized learning while protecting privacy and reducing communication overhead.