Docling, small models tackle Markdown docs

Tiny chat models usually fail on the hard part here, which is keeping tables, headings, and reading order intact. The better answer is a two-stage pipeline: extract structure first, then have a small model rewrite or summarize the cleaned text.

• –Docling already exports Markdown and is built for local, offline document processing, which makes it a good fit for sensitive docs.
• –Granite Docling is purpose-built for end-to-end document conversion, so it handles layout and structure better than a generic prompt.
• –For generation, 3B-9B class instruct models are the right range; sub-1B models are usually too brittle for consistent formatting.
• –On a 5070 Ti, the winning setup is likely throughput-first parsing plus a modest model, not one tiny model doing everything.