LoginSign Up
← Back to all sparks
D

Docling

AI-ASSISTANTS
Velocity5.0

Document parsing and conversion toolkit for AI pipelines

docling-project.github.io/docling ↗

Docling adds three more OCR model presets and keeps fixing format edge cases.

document-parsingocrvlmmodel-presetsformat-coverageopen-source
◆Current state
Docling keeps releasing every few days. Late October improved PDF quality, with broken text layers replaced by OCR and page orientation detection in RapidOCR, and hardened parsers against hostile files. 2.146.0 on November 3 adds presets for Unlimited-OCR 2, LightOnOCR-3 with grounded output and a guarded multi-pass Nemotron 3.5 parser, along with fixes across Markdown, DOCX, PPTX, XLSX, iWork and AsciiDoc.
◆Where it's heading
Docling acts as a hub for document vision models: every few releases add presets for a new OCR or VLM parser, from Nemotron-3.5-Parse in October to three more in November. Alongside, it widens inputs (ASR, ZIPs, iWork, AFP) and fixes long-tail format bugs, which suggests users run it on large, mixed document sets.
◆Prediction
More VLM and OCR presets will follow at the same pace, likely with more grounded output that ties text to page positions. Format-specific fixes for Office and iWork files will continue in each release.

◆Recent moves

  1. 1d ago

    New OCR presets: Unlimited-OCR 2, LightOnOCR-3, Nemotron 3.5

    Three new model presets, including grounded output for LightOnOCR-3 and a guarded multi-pass Nemotron 3.5 parser, continue Docling's role as a switchboard for document vision models. The rest is a long list of format fixes, from hanging Markdown parsing to images placed in Excel cells.

    View source ↗
  2. 6d ago

    RapidOCR page orientation and InputDocument.release()

    RapidOCR now detects page orientation and InputDocument.release() gives callers control over memory, with a long list of format fixes. Steady parsing quality work, including unloading page backends.

    View source ↗
  3. 7d ago

    OCR replaces broken PDF text layers; --heading-hierarchy option

    Broken PDF text layers are now replaced with OCR, and the heading style signal gets a --heading-hierarchy option. The text-layer fix addresses a common failure in real-world PDFs.

    View source ↗
  4. 8d ago

    Table token fixes for OCR word cells

    Fixes OCR word cells missing from table tokens, plus pptx and Tesseract details. A patch following 2.143.0.

    View source ↗
  5. 8d ago

    Mixed-format ZIP inputs and new ASR presets

    Mixed-format ZIP inputs, offline WhisperS2T presets, and transformers and MLX ASR presets widen what Docling takes in. Speech moves further toward being a first-class input.

    View source ↗
  6. 12d ago

    Size limits and parser bounds for untrusted documents

    Limits on decompressed archive members and decoded image pixels, bounded parsers, and a trusted-input requirement for Tectonic TikZ rendering. It is a hardening release that treats documents as potentially hostile.

    View source ↗