In multi-intent spoken language understanding (SLU), intent detection first comprehends and predicts the user’s intent, followed by slot filling, which retrieves the specific information needed to execute that intent. Current methods for multi-intent recognition have two main limitations: 1) They focus only on how intents guide slot filling, overlooking the bidirectional influences between the two tasks; 2) when using the detected intent information guides slot filling, the merging of multiple intents with each token can introduce unrelated intents, potentially interfering with slot predictions. This paper introduces a model integrating global interaction and bottleneck fusion for the combined detection of multiple intents and slot filling. Specifically, we adopt a global interactive module to promote mutual guidance between the two subtasks. The global interactive module consists of two attention mechanisms: intent-guided slot attention and slot-guided intent attention. Subsequently, we introduce a bottleneck fusion integrated into the multi-layer transformer. Bottleneck fusion explicitly fuses intent and slot features, extracting each slot’s most relevant intent information to avoid incorrect predictions. Experimental results from two public datasets demonstrate our model’s superior performance.