User-item interactions are typically driven by diverse and hierarchical intentions. Recent studies show that graph neural networks incorporating graph contrastive learning and disentangled representations have exhibited powerful performance by learning complex and diverse intents especially when the user behavior is usually inadequate in reality. While some of them show their effectiveness, existing works generally focus on modeling user intentions at a single level, lacking fine-grained and multi-level capture of intentions, which limits the model’s expressiveness in complex scenarios. Additionally, these methods fail to effectively control the information redundancy between different intentions when generating disentangled representations, thus affecting the independence and discriminative power of the representations. To address this, we propose a Multi-level Disentangled Contrastive Collaborative Filtering framework (MDCCF), which models user intentions at multiple levels and diversities through joint disentanglement and contrastive learning, generating more fine-grained and discriminative embedding representations. Specifically, MDCCF consists of two core modules: the Multi-level Disentangled Encoder, which extracts multi-level intent features from user-item interaction behaviors using convolution operations and multi-channel attention mechanisms, combined with graph convolutional networks to perceive intentions on the user-item interaction graph, generating disentangled embedding representations layer by layer; and the Contrastive Learning Module, which generates multi-view embeddings through a twin module, optimizing the quality of disentangled representations by applying orthogonal constraints to encourage the independence of different intentions while enhancing the model’s ability to capture multi-intent behaviors of users and items. The experimental results demonstrate that MDCCF outperforms existing comparative methods in terms of substantial improvement.