The team at the Institute of Software, Chinese Academy of Sciences recently open-sourced DeepPresenter, its second-generation intelligent agent system for slide creation. This system integrates presentation generation models with a complete agent sandbox environment into an integrated release, supporting one-click deployment on single consumer-grade graphics cards and standard terminals while remaining compatible with domestic computing power ecosystems, thereby further lowering the barrier to entry for AI-generated presentations. In terms of functional design, DeepPresenter establishes a shared agent environment incorporating over 20 tools, integrating capabilities such as Docker sandboxes, real-time arXiv literature retrieval, Python code execution, and data visualization to enhance professionalism in content generation and accuracy of data. Additionally, the system introduces a "generation-render-review-correct" visual closed-loop mechanism that performs screenshot checks and automatic optimization after generating each slide page, achieving adaptive layout adjustments. Regarding training methodology, the team constructed diverse task sets based on PersonaHub and arXiv datasets, introducing independent review mechanisms to conduct external evaluation and error correction on generated results. Ultimately, 802 high-quality samples were selected from 1,152 tasks for model training, covering both Chinese and English languages along with multiple complex constraint scenarios. During testing, DeepPresenter (9B version) achieved a score of 4.19 points on PPTEval evaluations, approaching GPT-5's performance level of 4.22 points and outperforming various mainstream solutions. Analysis indicates that this system achieves superior balance between computational cost and generation quality, while also supporting exportable editable pptx files for subsequent modifications and usage.