Abstract: In the signal detection of massive multiple-input multiple-output （MIMO） systems， deploying deep learning models on resource-constrained hardware typically requires optimizing storage overhead， computational complexity， and energy consumption. Model quantization provides a feasible approach to improve deployment efficiency. This paper proposes a mixed-precision model quantization method based on quantization-unit sensitivity. First， according to the position and functional role of each layer in the deep learning model， we partitioned the network into quantization units by grouping layers with similar structural characteristics； then， we transformed the hierarchical bit-allocation problem into a quantization-unit-level allocation problem to reduce the search space. Second， we designed a quantization-unit sensitivity metric that jointly accounts for detection performance and energy overhead and assigned bit widths based on sensitivity differences； within a limited bit-width range， we searched for an appropriate quantization configuration. Meanwhile， an energy-consumption model was constructed to include computation， weight transmission， and activation-value transmission， which is used to compute the sensitivity metric and evaluate quantization strategies. Simulation results showed that， compared with the full-precision model， fixed-precision quantization schemes， and conventional linear detection algorithms， the proposed method achieved a detection performance close to that of the full-precision model and outperformed fixed-precision quantization under the same total bit-width constraint and conventional linear detection. In addition， its energy consumption was significantly lower than that of the full-precision model.