Federated learning is a promising approach that omits the central learning of the collected data through allowing the participating nodes to learn locally their model, and eventually the submitted updates are aggregated to obtain a global trained model. The main goal of this process is maintaining the participant data privacy. Blockchain is supported by lots of federated learning applications in order to handle many issues such as the single point of failure, trust, integrity, privacy and security. On top of that, Blockchain scalability issue is one of the main problems that affects the efficiency of the federated leaning. A few works have been proposed to use Blockchain sharding, yet they have some shortcomings such as static sharding. Besides, the BC approaches marginalize the global model sharing or transfer process between the model owner and the task requester. In this work, we propose FLBCShard a new Blockchain sharding approach that uses proxy nodes and a dynamic shard formation in order to improve the system scalability and improving malicious nodes detection. In addition, for a better protection of the model ownership, we use IPFS and NFT based model sharing between and transfer. The proposed work is implemented using permissioned Hyperledger Fabric and evaluated using MNIST dataset.