PDF to Markdown benchmark: speed and reliability
Most "benchmarks" are synthetic. This one is real: every number below comes from production traffic over a two-week window, roughly 2,940 actual conversions. No cherry-picked sample, no lab conditions.
The numbers
Production traffic, 2026-06-10 to 2026-06-24. Processing time is measured server-side; it excludes any time spent waiting in the queue.
99.8%
conversion success rate (5 failures in ~2,945 attempts)
~2,940
real documents converted in the 15-day window
~14 s
median time for a short PDF (1 to 5 pages)
~4.8 s
per page, on the default engine, across all sizes
How long a conversion takes
Time scales with page count, as you would expect. These are the default engine's figures (the largest sample, 2,899 conversions). Median is the typical case; the 90th percentile shows the slow tail.
| Document size | Conversions | Median time | 90th percentile |
|---|---|---|---|
| 1 to 5 pages | 975 | 14 s | 42 s |
| 6 to 20 pages | 959 | 56 s | 3.1 min |
| 21 to 50 pages | 508 | 2.4 min | 7.2 min |
| 51+ pages | 448 | 8.2 min | 16.3 min |
A real spread of documents: nearly 1,000 short files, ~960 medium, and ~950 long (21+ pages). The figures are not skewed by only converting easy documents.
What happens at scale
Over the window, conversion succeeded 99.8% of the time. The handful of failures and partials are worth being honest about.
5 failures total
Out of ~2,945 attempts: four were timeouts on very large documents, one was an empty-output edge case.
~4% partial results
The longest documents can hit a soft time budget and return a flagged partial rather than nothing. The rest convert in full.
Near-zero retries
The default engine averaged 0.005 retries per job: conversions almost never needed a second attempt.
MinerU and Docling
The service runs two open-source engines. They show a clear, expected trade-off in this data.
MinerU – the workhorse
The default engine and the bulk of the sample (2,899 conversions, 99.8% success). It handles dense, multi-column and large documents robustly, at about 4.8 seconds per page. Every size bucket above is MinerU.
Docling – lighter and faster
On clean, smaller documents Docling is quicker, about 2.6 seconds per page (8.9 s median for 1 to 5 pages versus 14 s). Its sample here is small (41 conversions), so we report it as a speed signal on clean files, not a head-to-head verdict.
How this was measured, and what it does not cover
What we measured
What it does not cover
These are real-world figures on shared infrastructure, so absolute times depend on load. They are most useful as a guide to how conversion time scales with document size, and as evidence that conversion is reliable at scale.
Try it on your own PDF
Numbers are one thing; your document is another. Convert one free in the browser, or drive the same conversion from the API.
Common questions
How was this benchmark measured?
On real production traffic between 2026-06-10 and 2026-06-24: about 2,940 conversions, timed server-side as ready minus start. Success means the job reached the ready state. It is measured, not a synthetic test.
What is the success rate?
99.8%. There were 5 failures out of about 2,945 attempts in the window, the failures being timeouts on very large documents and one empty-output edge case.
How long does a conversion take?
It scales with document size. Median processing time is about 14 seconds for 1 to 5 pages, 56 seconds for 6 to 20 pages, 2.4 minutes for 21 to 50 pages, and 8.2 minutes for documents over 50 pages.
Does the benchmark measure accuracy?
No. This benchmark covers speed and reliability. Accuracy, meaning table, formula and OCR fidelity, needs a separate labelled evaluation and is a follow-up.
Which engine is faster?
Docling is lighter and faster on clean, small documents at about 2.6 seconds per page; MinerU is the default and handles dense, complex and large documents robustly at about 4.8 seconds per page. The MinerU sample is far larger, so docling's speed is reported with that caveat. See the converter roundup.
What does a partial result mean?
Very long documents can hit a soft time budget and return a flagged partial result; about 4% of jobs in this window did. The rest converted in full.