{"full_name":"datalab-to/marker","owner":"datalab-to","name":"marker","shard":"1","url":"https://github.com/datalab-to/marker","description":"Convert PDF to markdown + JSON quickly with high accuracy","homepage":"https://www.datalab.to","topics":[],"tag_pages":[],"language":"Python","license":"Apache-2.0","avatar":"https://avatars.githubusercontent.com/u/176631703?v=4","created_at":"2023-10-30T20:14:08Z","pushed_at":"2026-11-14T04:35:23Z","metrics":{"stars":40926,"forks":2954,"watchers":138,"issues":714,"prs":412,"releases":74,"issues_open":80,"issues_closed":634,"prs_open":1,"prs_closed":161,"prs_merged":250,"discussions":0},"growth":{"day":{"stars":16,"forks":0,"base_day":"2026-11-15"},"week":{"stars":102,"forks":-1,"base_day":"2026-11-09"}},"history":{"dates":["2026-11-09","2026-11-10","2026-11-11","2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[40824,40836,40860,40884,40887,40902,40910,40926],"forks":[2955,2956,2956,2956,2955,2955,2954,2954],"watchers":[139,139,139,138,138,138,138,138],"issues_open":[null,null,81,81,80,80,80,80],"prs_open":[null,null,61,11,1,2,1,1]},"boards":{},"same_owner":[{"full_name":"datalab-to/chandra","name":"chandra","shard":"1","boards":{}}],"issues":[{"number":985,"title":"[BUG: Breaking] OpenAI-compatible API (DeepSeek) fails with \"This response_format type is unavailable now\" when using `--use_llm`","comments":4,"created_at":"2026-02-03T03:05:21Z"},{"number":886,"title":"[BUG: Breaking] Marker cannot process files with signature lines","comments":4,"created_at":"2025-09-15T08:47:06Z"},{"number":1107,"title":"[Feature Request / Bug] Issues with mid-sentence image insertion, internal link escaping, and output file extension","comments":3,"created_at":"2026-09-14T07:04:11Z"},{"number":1065,"title":"[BUG: Breaking] Marker 2.0.0 fails on macOS after clean installation","comments":3,"created_at":"2026-07-21T17:54:14Z"},{"number":1013,"title":"[FEAT]Support offline custom directory artifacts for models&fonts","comments":3,"created_at":"2026-03-18T14:57:23Z"},{"number":988,"title":"[BUG: Output] Text lines dropped during PDF extraction (In vigore dal dates missing)","comments":3,"created_at":"2026-02-04T23:49:02Z"},{"number":1124,"title":"[BUG: Output] Tables that get no HTML lose their text (disable_ocr, or rejected table OCR)","comments":1,"created_at":"2026-10-19T14:07:49Z"},{"number":1066,"title":"[BUG: Breaking] MinerU Benchmark Uses Pipeline Instead of the Default Hybrid Medium Configuration","comments":2,"created_at":"2026-07-21T19:55:14Z"}],"discussions":[],"latest_release":{"name":"Marker v2.2.0 — 4 Conversion Modes, In-Process Indra ONNX Engine, Interactive Web Viewer & 88.9% on OmniParseBench","tag":"v2.2.0","published_at":"2026-11-12T17:45:53Z","body":"# Marker v2.2.0 — 4 Conversion Modes, In-Process Indra ONNX Engine, Interactive Web Viewer & 88.9% on OmniParseBench\n\nMarker `2.2.0` is our biggest architecture and usability upgrade since Marker 2.0:\n\n- **Up to 20× faster CPU conversion (`7–33 pages/s`) with zero external servers** via the new in-process **Indra ONNX** layout, reading order, and OCR engine (`fast` and `no_ocr` modes).\n- **88.9% on [OmniParseBench](https://huggingface.co/datasets/datalab-to/omni_parse_bench)** in `balanced` mode — routing clean digital pages through `pdftext` + Indra ONNX and routing tables, math, CJK/non-Latin scripts, and scanned pages to **Surya 2.5 VLM** (`86.7%` in pure `vllm` mode), outperforming Docling (`74.9%`) and PaddleOCR-VL (`84.4%`).\n- **New built-in Interactive Web Viewer (`marker_gui` & `marker_server`)** — replaces the legacy Streamlit app with a fast, self-contained browser UI featuring interactive SVG layout/line/reading-order overlays, stage timing bars, and `Preview` / `Markdown` / `Blocks` / `Lines` / `HTML` / `Chunks` / `JSON` inspection tabs.\n- **Agent-ready CLI (`--stdout`, `--skill`, `--install_skill`, `--recursive`, `--skip_existing`)** — pipe converted output directly to `stdout` or install the bundled AI agent skill (`SKILL.md`) for Claude Code, Cursor, and Antigravity in one command.\n\n> Need handwriting, complex forms, or managed VLM parsing at scale? Try **[Chandra on the Datalab API](https://www.datalab.to/?utm_source=gh-marker-release)** — get **5,000 free pages every month** when you sign up with a work email.\n\n---\n\n## 1. Four Conversion Modes (`--mode no_ocr | fast | balanced | vllm`)\n\nMarker now automatically picks the best mode for your hardware (`balanced` on CUDA GPUs, `fast` on CPU / Apple Silicon), or you can select any mode explicitly via `--mode`:\n\n| Mode | Speed (GH200 / CPU) | OmniParseBench | How It Works & When to Use |\n| :--- | ---: | ---: | :--- |\n| **`no_ocr`** | **32.7 pg/s** | 71.1% | Pure embedded PDF text layer (`pdftext`) + in-process Indra ONNX layout & reading order on CPU. Zero OCR calls. Ideal for born-digital PDFs when maximum CPU throughput is needed. |\n| **`fast`** | **7.0–33 pg/s** | 83.2% | In-process Indra ONNX layout, reading order, and ONNX line/page OCR on CPU/MPS/GPU. **Requires no external VLM server.** Default on CPU and Apple Silicon. |\n| **`balanced`** | **2.9 pg/s** | **88.9%** | Hybrid routing: `pdftext` + Indra ONNX on clean Latin prose pages; routes tables, math/equations, CJK/non-Latin scripts, and scanned/garbled pages to **Surya 2.5 VLM**. Default on CUDA GPUs. |\n| **`vllm`** | **2.5 pg/s** | **86.7%** | Full-page **Surya 2.5 VLM** OCR on every page, skipping `pdftext` and Indra entirely. Best for degraded scans, dense math, and complex multi-script layouts. |\n\n### OmniParseBench Comparison (838 pages, 15,834 unit tests)\n\n| Parser | Headline | Text | Tables | Math | Layout | Order | Pages/s |\n| :--- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |\n| **Marker 2.2 (`balanced`)** | **88.9** | **93.5** | **89.0** | 75.7 | **92.8** | **92.4** | 2.9 |\n| **Marker 2.2 (`vllm`)** | 86.7 | 92.8 | 88.2 | **78.0** | 89.8 | 91.6 | 2.5 |\n| PaddleOCR-VL 1.5 | 84.4 | 92.1 | 83.1 | 76.2 | 82.9 | 89.2 | 1.1 |\n| **Marker 2.2 (`fast` - CPU ONNX)** | 83.2 | 90.2 | 81.2 | 66.2 | 91.2 | 90.5 | **7.0** |\n| Docling | 74.9 | 84.9 | 78.2 | 35.1 | 81.4 | 85.4 | 1.4 |\n| **Marker 2.2 (`no_ocr` - CPU)** | 71.1 | 79.1 | 69.2 | 47.0 | 88.2 | 87.0 | **32.7** |\n\n---\n\n## 2. Built-In Interactive Web Viewer (`marker_gui` & `marker_server`)\n\nRun `marker_gui --port 8000` (or `marker_server --port 8000` for the FastAPI REST server + viewer + `/docs`) to inspect documents page by page in your browser:\n\n- Drag & drop PDFs, images, `DOCX`, `PPTX`, `XLSX`, `HTML`, or `EPUB` files.\n- Switch live between `Auto`, `No OCR`, `Fast`, `Balanced`, and `vLLM` modes.\n- Interactive SVG overlays for **layout blocks**, **block labels**, **reading order**, and **text lines** with bidirectional hover highlighting.\n- Inspe","is_prerelease":false},"readme":true,"readme_zh":false,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}