<p>Traffic flow prediction is an essential area of intelligent transportation systems with broad practical applications. Recently, Large Language Models (LLMs) that integrate traffic data with textual prompts have achieved remarkable performance improvements in traffic flow prediction. However, we observed that current LLM-based methods struggle to effectively align numerical traffic sequences and textual prompts during cross-modal fusion, leading to inter-modal interference that limits further performance gains. To address this issue, we propose TrafficMCA, an LLM-empowered framework for traffic flow prediction via mutual cross-attention. Specifically, we design a dual-modality encoding module comprising two branches: the traffic flow encoding branch extracts fundamental spatio-temporal features from traffic data by integrating timestep, hour-of-day, and spatial embeddings, while the prompt text encoding branch leverages a pre-trained encoder to extract rich semantic features from textual prompts. To enhance cross-modal fusion, we introduce a Mutual Cross-Attention (MCA) mechanism that explicitly captures complementary information between the two modalities, enabling collaborative guidance and bidirectional enhancement of features. As another key design, we adopt Low-Rank Adaptation (LoRA) to fine-tune the pre-trained LLM backbone in TrafficMCA, which significantly reduces computational overhead while effectively maintaining predictive performance. Extensive experiments demonstrate that TrafficMCA outperforms 15 state-of-the-art methods and exhibits strong generalization capabilities in few-shot and zero-shot scenarios.</p>