Recent advances in multimodal large language models (MLLMs) have delivered remarkable gains in multimodal understanding and reasoning, yet their prohibitive computational and memory costs remain a major barrier to real-world deployment. We present Mamba-based Multimodal Vision-Language Quantization (Mamba-MVQ), a compact and efficient multimodal inference framework that unifies the linear-complexity Mamba architecture with alignment-aware mixed-precision quantization. The proposed method assigns higher precision to semantically critical cross-modal modules, such as visual projection layers, while aggressively quantizing less sensitive components, including the linear layers in Mamba blocks, to maximize compression without severely compromising performance. To mitigate cross-modal semantic degradation introduced by quantization, we further introduce an alignment-preserving calibration objective based on vision-language contrastive learning. Combined with dynamic token merging, Mamba-MVQ reduces inference cost from both architectural and token-level perspectives. Extensive experiments on OpenCompass and MMBench show that Mamba-MVQ reduces model memory to 23.8% of the full-precision baseline with only a 1.7% average accuracy drop, demonstrating an effective trade-off between efficiency and multimodal capability.
