hanji-dev/hanji-parse-4b

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Hanji Parse 4B is a 4 billion parameter vision-language model developed by hanji-dev, fine-tuned from Qwen3-VL-4B-Instruct. It specializes in document parsing, extracting layout-grounded content blocks from page images and emitting them as a JSON array. This model excels at segmenting semantic sections like paragraphs, headings, key-value panels, and tables, providing bounding boxes and transcribed text for each.

Loading preview...

Hanji Parse 4B: Document Parsing VL Model

Hanji Parse 4B is a specialized 4 billion parameter vision-language model, fine-tuned from Qwen3-VL-4B-Instruct, designed for robust document parsing. Its core function is to take a page image and output a JSON array of layout-grounded content blocks. Each block represents a semantic section of the page, such as a paragraph, a heading with its associated content, a key-value panel, or an entire table, complete with a bounding box and transcribed text.

Key Capabilities and Output Contract

  • JSON-only Output: The model consistently returns a JSON array of records, each containing bbox_2d (normalized 0-1000 page coordinates) and text_content.
  • Semantic Block Grouping: It groups content into semantic sections (typically 5-30 per page), rather than individual lines, cells, or fields, ensuring meaningful data extraction.
  • Table Handling: Tables are transcribed as a single block containing GitHub-Flavored Markdown, preserving cell content and structure.
  • Image Recognition: Non-text graphics like photos, charts, and signatures are identified as image blocks with text_content set to "".
  • Checkbox Transcription: Checkboxes are transcribed inline as [x] for marked and [ ] for unmarked.
  • Meaning-Focused Transcription: The model prioritizes transcribing meaning over layout glyphs, omitting decorative elements and blank lines, and merging labels with values.

Usage and Optimization

  • Image Preprocessing: Requires specific image downscaling (max 2MP) and dimension flooring to multiples of 32 for optimal performance.
  • Strict Prompt Adherence: Users must use an exact, predefined prompt to guide the model's parsing behavior.
  • Decoding Recommendations: Recommends greedy decoding (temperature 0.0, top_p 1.0) and strongly advises JSON-schema-constrained decoding for reliable output.
  • Performance: Optimized for serving with SGLang, leveraging N-gram speculative decoding for significant speedups on its structured JSON output.

Finetunes
1 models