In recent years, diffusion models have emerged as a novel paradigm in generative modeling, achieving breakthroughs in image, audio-video, and text generation tasks. However, in the domain of text generation, existing continuous diffusion-based models struggle to efficiently capture temporal dependencies in long sequences and exhibit limited robustness in complex noisy environments, which significantly affects the quality and diversity of the generated text. To address these challenges, we propose MM-Diffusion, a novel continuous diffusion-based text generation model. MM-Diffusion incorporates a parallel architecture that combines Mamba and Transformer during the reverse denoising process, leveraging Mamba’s efficiency in long-sequence modeling and the Transformer’s ability to capture complex semantics, thereby facilitating effective modeling of long-text temporal dependencies. Furthermore, we introduce an adaptive noise adjustment module based on a mixture-of-experts framework. The dynamic routing mechanism within this module adaptively activates a variable number of expert networks based on token complexity, enabling fine-grained noise control and enhancing text diversity. Experimental results on four benchmark datasets demonstrate that MM-Diffusion achieves effective improvements across multiple performance metric.