PaliGemma 2

PaliGemma 2 pairs an upgraded vision encoder with a compact Gemma decoder to “look, read, and reason.” It ingests one or more images alongside a prompt and produces grounded, step-by-step text—captions, answers, summaries, or structured outputs (Markdown/JSON). Compared with the original PaliGemma, it improves layout-aware OCR, table/chart interpretation, and screenshot/UI analysis, and handles higher-resolution inputs via tiling/cropping strategies for dense documents.
For builders, it’s instruction-tuned for reliable formatting, supports function/tool calling for agent workflows (e.g., crop → read → reason), and integrates cleanly with RAG so answers can cite or reference specific regions. It’s lightweight enough to run on a single modern GPU, with 8/4-bit quantization and LoRA/full fine-tuning options to adapt to domains (invoices, forms, dashboards, manuals). Typical uses include enterprise document automation and extraction, analytics over charts/dashboards, accessibility (image descriptions), and developer assistants that reason directly from screenshots—bringing practical, efficient visual understanding to the Gemma ecosystem.