Gemma 4 OCR demo tests llama.cpp stack

This is the right kind of demo for Gemma 4: not benchmark theater, but a workflow that reveals whether the model and runtime are both ready for real documents.

• –Google explicitly positions Gemma 4 as strong at OCR and chart understanding, so this use case maps directly to the launch claims.
• –llama.cpp support is the real gating factor for local users; model quality does not matter if template, tokenizer, or multimodal plumbing is brittle.
• –If this works cleanly, Gemma 4 becomes more than a chat model: it’s a viable local document-extraction and vision pipeline for offline workflows.
• –The community already seems focused on runtime fixes and backend compatibility, which makes real OCR tests more informative than synthetic leaderboards.