Network slicing is a key technology to allow resource sharing among heterogeneous operators/ services, which achieves QoS isolation for service provisioning in future communication networks. In this article, a comprehensive hierarchical soft-slicing framework is proposed to enable software-defined radio access networks supporting differentiated services with diverse QoS requirements. The proposed framework consists of network-level slicing and gNodeB-level slicing. In the network level, radio RBs are pre-allocated to each gNodeB in a large time scale, while in the gNodeB level, the pre-allocated RBs are dynamically scheduled to the services in response to the small time scale (mini-slot-level) RB request variations. The proposed framework allows the accommodation of time-varying traffic loads of differentiated services over multiple gNodeBs, while enabling dynamic inter-gNodeB RB sharing to increase the resource multiplexing gain. A case study is presented to demonstrate the effectiveness of the proposed framework, followed by a discussion on open research issues.