Uni-FinLLM is a unified multimodal large language model designed to jointly handle micro-level stock prediction and macro-level systemic risk assessment. It employs a shared Transformer backbone with modular task heads, enabling the integration of diverse financial data types—text, numerical time series, fundamentals, and visual data—through cross-modal attention and multi-task optimization. This architecture allows the model to learn coherent representations across different scales, from individual stock behavior to systemic vulnerabilities. The model is evaluated on three tasks: stock directional forecasting, credit-risk assessment, and systemic risk early-warning detection. Results show significant improvements over baseline methods, with stock accuracy rising from 61.7% to 67.4%, credit-risk accuracy from 79.6% to 84.1%, and macro early-warning accuracy reaching 82.3%. The work demonstrates that a unified multimodal LLM can effectively capture cross-scale dependencies and serve as a scalable decision-support engine for financial institutions and regulators.