Docling vs the alternatives — measured on the same PDFs

Which tool fits which job#

Speed on the same PDF#

All tools ran on the same 9-page born-digital paper (the Docling technical report) on a 4-vCPU machine without a GPU. Docling's two bars show the default pipeline and the same pipeline with OCR off — OCR of the paper's figures is what costs the time. Marker 1.10.2 took 90.7 s/page here — its models are built for GPUs, so expect very different numbers on one. Unstructured's default strategy reads the text layer (0.36 s/page); its hi_res mode with table detection, which we used for the table test, took 10.0 s/page.

Scanned pages#

Four pages of the same paper rasterised at 150 DPI with slight blur and skew, so there is no text layer. The score is the share of unique words (four letters or more) of the original text that appear in each output. PyMuPDF4LLM and Unstructured OCR through Tesseract, which must be installed separately (without it, PyMuPDF4LLM returned no text at all). Unstructured's text-only output scored highest; Docling's Markdown deliberately leaves out page headers, footers and the arXiv margin stamp, which costs a few points on this word-count metric. Docling's own engines are compared in the OCR guide — in our 2.148.0 re-run with Tesseract as its engine, Docling reached 96.1% at 5.5 s/page. This chart is from our 27 October run with Docling 2.143.1, on the scan copy all five tools were tested with.

The same table, five outputs#

Table 1 of the paper has a two-level header and rows where each cell holds two values (4 and 16 threads). No horizontal rules separate the sub-rows — a classic trap. Here is exactly what each tool returned (trimmed to the table):

| CPU                                         | Thread budget   | native backend - TTS   | native backend - Pages/s   | native backend - Mem   | pypdfium backend - TTS   | pypdfium backend - Pages/s   | pypdfium backend - Mem   |
|---------------------------------------------|-----------------|------------------------|----------------------------|------------------------|--------------------------|------------------------------|--------------------------|
| Apple M3 Max                                | 4 16            | 177 s 167 s            | 1.27 1.34                  | 6.20 GB                | 103 s 92 s               | 2.18 2.45                    | 2.56 GB                  |
| (16 cores) Intel(R) Xeon E5-2690 (16 cores) | 4 16            | 375 s 244 s            | 0.60 0.92                  | 6.16 GB                | 239 s 143 s              | 0.94 1.57                    | 2.42 GB                  |

|CPU|Thread<br>budet|n<br>|ative back<br>|end<br>|pypi<br>|dfium bac<br>|i kend<br>|
|---|---|---|---|---|---|---|---|
||g|TTS|Pages/s|Mem|TTS|Pages/s|Mem|
|Apple M3 Max<br>(16 cores)|4<br>16|177 s<br>167 s|1.27<br>1.34|6.20 GB|103 s<br>92 s|2.18<br>2.45|2.56 GB|
|Intel(R) Xeon<br>E5-2690<br>(16 cores)|4<br>16|375 s<br>244 s|0.60<br>0.92|6.16 GB|239 s<br>143 s|0.94<br>1.57|2.42 GB|

Thread nativebackend pypdfiumbackend
CPU
budget
TTS Pages/s Mem TTS Pages/s Mem
AppleM3Max 4 177s 1.27 103s 2.18
6.20GB 2.56GB
(16cores) 16 167s 1.34 92s 2.45

| CPU                                    | Thread  |                | native backend |         | pypdfium backend |              |         |
|----------------------------------------|---------|----------------|----------------|---------|------------------|--------------|---------|
|                                        | budget  | TTS            | Pages/s        | Mem     | TTS              | Pages/s      | Mem     |
| Apple M3 Max<br>(16 cores)             | 4<br>16 | 177 s<br>167 s | 1.27<br>1.34   | 6.20 GB | 103 s<br>92 s    | 2.18<br>2.45 | 2.56 GB |
| Intel(R) Xeon<br>E5-2690<br>(16 cores) | 4<br>16 | 375 s<br>244 s | 0.60<br>0.92   | 6.16 GB | 239 s<br>143 s   | 0.94<br>1.57 | 2.42 GB |

<table><thead><tr><th rowspan="2">CPU</th><th rowspan="2">Thread budget</th><th colspan="3">native backend</th><th colspan="3">pypdfium backend</th></tr>
<tr><th>TTS</th><th>Pages/s</th><th>Mem</th><th>TTS</th><th>Pages/s</th><th>Mem</th></tr></thead>
<tbody><tr><td>Intel(R) Xeon</td><td>4</td><td>375s</td><td>0.60</td><td></td><td>239s</td><td>0.94</td><td></td></tr>
<tr><td>E5-2690</td><td>16 =.</td><td>244s</td><td>0.92,</td><td>FOGB</td><td>jy35</td><td>p57</td><td>742 GB</td></tr></tbody></table>

Licences: the row that decides most projects#

Read from each package's PyPI metadata and licence file on 27 October 2026. This is not legal advice.

Features at a glance#

Format lists come from each package's own description or source tree in the versions above; “—” means we did not find it in the package, not that it is impossible.

What we could not test#

• Marker 2.0.0 runs its models through a local inference server (vLLM on NVIDIA GPUs, llama.cpp elsewhere). We could not get that server running in our sandbox, so we benchmarked Marker 1.10.2, the last release with the earlier architecture. Marker 2's speed and quality may differ substantially.
• MinerU 4.0.12 installed, but its bundled llama.cpp binary crashed on our machine (free(): invalid pointer) before converting a page. We report its licence and formats only.
• GPUs. Every number here is CPU-only. Model-heavy tools (Docling with OCR, Marker, MinerU) gain the most from a GPU; text-layer tools gain little.

How we tested#

• Each tool in its own Python 3.13 virtual environment, default settings, one warm-up conversion, then timed runs in the same process (medians of three runs; Marker, at up to several minutes per page, one timed run after a warm-up on the born-digital PDF and a single run with models already loaded on the scans).
• Hardware: 4 vCPU x86_64, 15 GB RAM, no GPU. Documents: the Docling technical report (arXiv 2408.09869, 9 pages) and a rasterised copy of its first four pages.
• We publish raw outputs on request: hello@docling.org. Tool maintainers who think a setting was unfair: tell us and we will re-run.

Frequently asked questions

Sources

• Our benchmark method and hardware
• Marker README: licence and commercial usage (marker-pdf 2.0.0)
• MinerU Open Source License (mineru 4.0.12)
• PyMuPDF licensing (AGPL-3.0 or Artifex commercial)
• MarkItDown on GitHub (MIT)
• Unstructured open-source library (Apache-2.0)