Federated learning (FL) is a distributed machine learning paradigm that allows participants to collaboratively train a neural network model. Its unique data-sharing approach has greatly facilitated the development of Artificial Intelligence of Things (AIoT). However, federated learning suffers from the single-point failure problem, which cannot provide a stable and reliable interaction environment for IoT devices. An effective solution is to use a decentralized blockchain network to replace the central server for federated learning, but this inevitably increases the communication delay of the neural network model within the blockchain network. Furthermore, the transmission of neural network models among IoT devices also faces the issue of long waiting times due to resource-constrained devices. To address these issues, this paper introduces an adaptive compression scheme (BFLD) based on the blockchain-based federated learning model. By deploying smart contracts on the blockchain network, BFLD generates smaller network models for resource-constrained IoT devices to replace the global model, reducing the communication latency of the network model in both the blockchain network and the IoT devices, thus lowering the overall communication time overhead of the system. Experimental results indicate that the BFLD architecture can significantly reduce the communication time overhead of each training round in federated learning under the blockchain network while ensuring the performance of the network model.