Most document parsers force a choice: stitch together separate detectors, OCR engines, and table models that drift out of sync, or hand the whole page to a large general-purpose VLM and eat the latency. Dolphin takes a third path—analyze the layout first, then let the discovered elements prompt the parsing of their own content.
Key Capabilities
- Analyze-then-parse, two stages. Stage 1 emits a sequence of layout elements in natural reading order; Stage 2 treats each element as an anchor paired with a task-specific prompt. So you are no longer bottlenecked by reading a whole page top-to-bottom—elements are parsed in parallel.
- Heterogeneous anchor prompting. Text, tables, formulas, figures, and code blocks each get a different prompt, so one model handles intertwined content instead of a pipeline of one model per element type that has to be kept in sync.
- Built on scale. Trained on 30M+ samples spanning multi-granularity tasks, it works at both page level and element level, and handles multi-page PDFs.
- Lightweight by design. It is far smaller than a general VLM, which is where the throughput advantage comes from rather than from raw model size.
Who It's For
Great fit if you need to turn messy real-world documents—phone photos, scans, multi-column PDFs—into structured Markdown or JSON with tables and formulas preserved, and you care about throughput over a do-everything model. Look elsewhere if you want a general-purpose VLM for visual Q&A or reasoning over images: Dolphin is specialized for parsing document structure, not conversational understanding. The repo now also ships an enhanced v2 checkpoint aimed at photographed documents, so check which checkpoint matches your inputs before committing.
