Abstract: To address the limitations of single-modal data and poor interpretability in traditional indicator diagram methods-based liquid production prediction for coalbed methane production wells, this study proposes a Qwen large language model-based indicator diagram prediction model for liquid production. A unified multimodal data processing approach is employed: indicator diagram image data are embedded as images, while such engineering parameters as stroke frequency, pump diameter and pump depth are embedded as texts. Leveraging Qwen's multimodal fusion mechanism, adaptive feature fusion is achieved through the large model's cross-modal attention layer. This establishes an interactive mode where multimodal data are taken as inputs and liquid production predictions as outputs along with interpretable fault diagnosis reports. In terms of prompt design, mass conservation equations are transformed into natural language prompt templates, and through prompt fine-tuning, the large model is guided to generate prediction results compliant with engineering principles. In testing in a specific block, the model demonstrated liquid production prediction errors below 7.85% across varying pump diameters (32–56 mm) and complex operating conditions such as insufficient fluid supply and wax deposition. This provides efficient and reliable technical support for the intelligent management of coalbed methane production wells.