Traffic flow prediction is essential for the advancement of intelligent transportation systems. However, existing models encounter significant challenges. For one thing, prevailing traffic graph convolutional networks (GCNs) rely on layer stacking to implicitly learn multi-hop spatial relationships, resulting in low interpretability and limited adaptability to dynamic, time-varying traffic conditions. For another, the multi-head attention (MHA) mechanism used in traffic field may contain redundant representations of traffic patterns in certain scenarios, resulting in unsatisfactory performance. To this end, we propose a multi-hop aware graph convolutional network and collaborative Transformer for traffic flow prediction (HopFormer). Specifically, we propose a multi-hop aware graph convolutional network (HAGCN) that explicitly and adaptively adjusts multi-hop influence weights in response to time-varying traffic conditions, while enhancing model interpretability. Moreover, we introduce a collaborative Transformer, where attention heads actively exchange spatial-temporal features through the compressor and restorer, effectively reducing inter-head redundant representations, improving complementary information sharing while preserving distinctive traffic pattern diversity. Experimental evaluations on three datasets show that our proposed HopFormer consistently outperforms state-of-the-art methods across three evaluation metrics, validating its effectiveness in improving traffic flow prediction.