The third part of the guide on building a local AI stack focuses on organizing model files and creating a static HTML cheat sheet. This tool allows users to quickly identify the best model for specific tasks like coding, roleplay, or OCR without needing to consult external documentation or configuration files.
The system uses a structured folder approach where standard models are stored in a flat directory, while multimodal models are placed in subfolders containing mmproj files. The static HTML page, which functions offline via a browser, includes a task-based selection table, model cards with pros and cons, and a copy-to-clipboard feature for launch commands.
The guide recommends using the Q4_K_M quantization level as an optimal balance between performance and size for CPU-based execution. Key models suggested for covering 90% of use cases include Qwen3.5-9B for general tasks, OxCoder-9B for coding, Gemma-4-E4B for roleplay, and Qwen3VL-4B for OCR and vision tasks.
To add a new model to the interface, users simply update the MODELS array in the script with the model's name, size, task tags, and description. The system automatically generates the corresponding card and filter buttons, ensuring the interface remains scalable and easy to maintain as the local library grows.
