Abstract: In the VUCA era characterized by volatile global economic policies, corporate financial crises exhibit new features of nonlinear transmission and multi-source risk coupling, rendering traditional early warning models obsolete. To address this, we propose a financial crisis early warning method based on the “data - text - classification" transformation path, leveraging the powerful language understanding and text processing capabilities of DeepSeek. This method converts structured financial data containing Z-score and 13 financial indicators into smooth enterprise data description, which is used to fine-tune the performance of pre-training model and realize accurate identification of financial crisis. Using financial data from 147 publicly listed food manufacturing enterprises spanning the years 2002 to 2023, experimental results show that the pre-trained model fine-tuned with DeepSeek-generated text achieves over 95% accuracy, macro-F1 score, and macro-average recall, significantly outperforming traditional models such as Random Forest and DeepFM. This study highlights the effectiveness of the DeepSeek model for enterprise financial crisis early warning, offering both theoretical contributions and practical insights for advancing AI-driven financial risk management.