Existing rolling bearing fault diagnosis methods often face challenges of high computational complexity and limited multimodal fusion efficiency, particularly in multimodal scenarios. To address these limitations, we propose Mamba TFVisionChaos, a lightweight multimodal fusion framework for rolling bearing fault diagnosis. Our method (1) employs a multi-scale Mamba-based temporal module that reduces computational complexity from $\mathcal{O}(L^2)$ to $\mathcal{O}(L)$ while capturing deep sequential patterns in vibration signals, (2) integrates an efficient image processing branch for extracting hierarchical features from time-frequency representations, and (3) introduces a chaos-enhanced detection module for early fault identification through phase space reconstruction. This design enables the model to achieve superior performance with exceptional parameter efficiency in multimodal data fusion scenarios. Extensive experiments on four benchmark datasets demonstrate the effectiveness of our approach. To ensure statistical reliability, we repeat each experiment five times under different random seeds and report mean accuracy (Mean $\pm$ Std): 99.90\% $\pm$ 0.23\% on CWRU, 100.00\% $\pm$ 0.00\% on XJTU-SY, 99.07\% $\pm$ 0.30\% on JNU, and 97.40\% $\pm$ 0.70\% on HIT. Across datasets, the model achieves AUC in the range of 0.9984--1.0000. The deployed model contains about 334K parameters ($\approx$1.28 MB, 0.231 GFLOPs); average inference latency is 82.6--107.5 ms per batch (batch size 64), and peak GPU memory is about 832 MB, providing a practical solution for real-time rolling bearing fault diagnosis in industrial environments.