With the advent of Industry 4.0/5.0, bearing fault diagnosis is crucial for intelligent manufacturing. This paper proposes DM²-Fuse, an interpretable multimodal framework based on large–small model collaboration. It integrates lightweight Mamba variants—MambaNAS (for raw vibration signals) and MambaDES (for time–frequency representations)—into MambaBearing to extract multimodal features. These are input to an LLM for diagnostic decisions and structured reports. Experiments on CWRU, JNU, and SEU show that the multimodal MambaBearing model achieves 99.41%–100% accuracy with only 146.4K parameters. MambaNAS achieves 40,009 samples/s encoder throughput (1.17× WDCNN), while MambaDES is 11.7× faster than ResNet18. Ablations confirm NAS and separable convolutions contribute +4.8% and +1.3% accuracy, respectively. An adaptive routing mechanism reduces average LLM latency by 86%, and extensive noise and cross-domain experiments validate practical robustness. The framework reduces computation, maintains high accuracy, and enhances interpretability for industrial deployment.