Discrete semantic IDs generated by quantization techniques are becoming essential for enabling efficient and effective large-scale recommendation systems. However, prevailing methods based on Residual Quantization often suffer from two critical limitations: 1) semantic degradation, where the core meaning of an item is progressively lost in deeper quantization layers, and 2) the hourglass phenomenon, where an imbalanced codebook distribution leads to information bottlenecks and restricts the model’s representational capacity. To address these issues, we propose Dense Residual Quantization, a novel framework that preserves semantic integrity throughout the quantization process. The core innovation of DenseRQ lies in reincorporating the original content embedding at each quantization layer by concatenating it with the corresponding residual vector. This approach anchors fine-grained detail learning to the item’s core semantics, thereby preventing semantic degradation and enabling nuanced item differentiation. Extensive experiments on public and large-scale industrial datasets demonstrate that DenseRQ significantly outperforms state-of-the-art methods in downstream recommendation tasks. Furthermore, online A/B tests validates the practical effectiveness of DenseRQ, yielding substantial improvements in key business metrics. In-depth analysis also reveals that DenseRQ generates a more balanced and expressive codebook distribution, effectively mitigating the aforementioned bottlenecks.