Datalab’s Marker 2 vs MinerU, Docling and LiteParse: 76.0 on olmOCR-bench at 5× MinerU’s Throughput

Datalab released Marker 2a complete rewrite of its open source document conversion pipeline. Markup converts PDF, image, PPTX, DOCX, XLSX, HTML, and EPUB files to markdown, JSON, HTML, or snippets. The Datalab team has rebuilt almost three components that were posted months ago: Surya OCR 2a 20M-param fast structure modeland reconstruction pdf document that’s 3× faster than before.

The main result comes from olmOCR-bench, a third-party benchmark from Allen AI. Marker 2’s balanced mode scores 76.0% overall and 83.5% for digital PDFs. It maintains 2.9 pages per second on a single B200 GPU. That’s more than 5× the output of the MinerU pipeline backend, which scores 72.7% at 0.54 pages per second. Docling scores 50.3% at 2.1 pages per second on the same harness.

What’s New in Marker 2

Tag 2 exposes three conversion methods instead of one:

• balanced — Surya VLM handles the layout, and the entire page is re-OCR’d whenever the embedded text is bad. Very high quality, best for GPU. 76.0% olmOCR-bench.
• fast — lightweight rf-detr/onnx structure detector and pdf document, low-power, surgical VLM implementation. 66.6%and very cheap.
• -disable_ocr – pure text layer output, no VLM calls at all. It runs entirely on the CPU. 43.6%23.7 pg/s.

Mode now device-by-default: balanced on GPU, faster on CPU/MPS, stronger in -mode. Full CPU support is the second design change. fast -disable_ocr requires no GPU and no inference server, and the 20M layout model still reads columns, tables and headers on the CPU.

The third change is the construction, and it is the one that produces the output numbers. Multiple sub-CPU workers share a single Surya inference server. The parent process budget VLM is consistent across, so power consumption scales with server capacity rather than per-process VRAM. Datalab reports that balanced mode supports ~2.9 pg/s against an average of ~0.3 pg/s for a single stream on the same hardware.

Breaking changes should be flagged before upgrading. Python 3.10+ is now required. Packaging moved from Poetry to uvby hatching as a build backend, although pip install marker-pdf is unchanged. The converter-extraction converter and extractors were removed; Datalab points users to the hosted API or -use_llm workflow instead.

Comparison

The benchmark score is olmOCR-bench from Ai2: 1,403 PDFs with approximately 8,400 tests/unit failures including figure rendering, table structure, reading order, headers and footers, and old scans. The overall score is the grand average across all 8 categories, combined with the official olmOCR bench tester. Work continues simultaneously pg/s for a single B200 host, not a single stream delay.

A note of appearance. olmOCR-bench is a third-party benchmark from Ai2, but all scores and output values ​​below are from Datalab’s runs. All are reproduced with an open harness on Marker’s site, sending competing MinerU, Docling and LiteParse runners alongside Marker’s own.

These numbers also reflect a benchmark document composition measured on a single hardware setup, so results on your documents may vary. Teams evaluating these programs must use a harness against their own corpus, the only way to know how they rank in the literature they are considering.

Marker 2 vs MinerU

The backend of the MinerU pipeline is very similar in structure. Both read the PDF text layer and optionally OCR. In total scores, the Balanced Marker leads at 76.0 to 72.7. In the born digital records the two are successfully tied: 83.5 against 83.3.

A separation is an exit. The benchmark supports 2.9 pg/s against MinerU’s 0.54 pg/s, a 5.4× gap in the top score. The fastest marker supports 7.4 pg/s, about 13.7× the MinerU pipeline rate, but it scored 6.1 points less than MinerU to do it.

MinerU also sends a Conclusion of VLMDatalab says it scores higher than its pipeline backend. That background is a full-page VLM method and is not in this table. AI teams testing MinerU should evaluate that approach separately.

Marker 2 vs Docling

Docking is the widest curve between GPU pipelines. The average score leads 76.0 to 50.3 overall and 83.5 to 64.0 in the digital birth, while also running quickly: 2.9 pg/s against 2.1 pg/s. Datalab Docling notes applied to its automation line, which uses a text overlay on native digital pages and OCR on image regions.

Docling’s counterweight is dominance and breadth of format, not accuracy. The codebase i MIT Licenseoriginated from IBM Research, and is hosted as a project at LF AI & Data Foundation. Its input list also extends past documents into audio and email formats.

Tag 2 vs LiteParse

LiteParse, from the LlamaIndex team, is a Rust document parser. It does not compete on the same axis. On CPU it scores 22.4 overall and 20.4 with OCR off, compared to Marker’s CPU-only 43.6.

But LiteParse with OCR blocked the reports 1721 pg/s – about 73 × Marker CPU mode, which is a tradeoff. Marker’s fast –disable_ocr uses a 20M layout model on the CPU and still returns the layout, hence doubling the empty dump points. LiteParse has no structure model and wraps around anything arbitrary.

Mark 2 compared to a full-page VLM section

The Datalab team emphasizes that Marker is designed as a pipeline rather than a VLM, clarifying that these are separate tools. In this experiment, the hosts Chandra 2 it gets a score of 85.8, while Gemini Flash 3.5 with an API score of 76.4. Chandra’s Datalab repository also lists OlmOCR 2 for Ai2 at 82.4 and dots.ocr 1.5 at 83.9 within a separate table. For scanning, math-heavy pages, and achieving high precision, the VLM class remains the best of all pipelines on the list.

The balanced markup mode narrows the performance gap to just 0.4 points behind Gemini Flash 3.5 overall, and even outperforms native digital documents by a margin of 83.5 to 79.1—without requiring a per-page API call.

The behavior of each category

The mode you choose changes the failure profile, not just the score. Each row is the olMOCR bench class, with scores for all three methods. Statistics is a sharp edge: fast mode reads statistics from the PDF text layer instead of VLM-OCRing them, so arXiv’s statistics drop from 83.9 to 23.4, and -disable_ocr scores 0.0 there by design. Apart from the two statistical categories, the old scan is the weakest classification of all modes, coming out at 43.2.

License

This is where the four systems differ the most from sales teams:

• Mark: code is Apache 2.0. Model weights use the modified AI Pubs OpenRAIL-M license — free for research, personal use, and startups under $5M in revenue/revenue. Additionally, commercial use of weights requires a paid license.
• MinerU: now under the MinerU Open Source License, based on Apache 2.0 with additional terms. A separate marketing license is required above 100M MAU or $20M in monthly revenue, and online services built upon it must disclose that fact.
• Docling: MIT, whose model licenses are tracked separately from their original packages.
• LiteParse: open source, from run-llama, with LlamaParse positioned as a paid cloud solution for robust documents.

Use Case- Comparison

Score alone does not select a tool. Corpus type, hardware, licensing band and output format determine that. Try the interactive selector below to filter ten use cases by restriction and tool, and see which filter fits your use case.

Key Takeaways

• Marker 2 scored 76.0% on the olmOCR-benchmark at 2.9 pg/s — more than 5× the output of the MinerU pipeline.
• It beats Docling on both axes simultaneously: 76.0% versus 50.3%, and 2.9 pg/s versus 2.1 pg/s.
• LiteParse trades for speed — 1721 pg/s with OCR off, but 20.4% versus Marker’s 43.6% on CPU.
• Fast mode with -disable_ocr runs entirely on CPU, no inference server, at 23.7 pg/s.
• Licensing divides the field: To be documented by MIT, MinerU remains free with a monthly revenue of $20M, and Marker weights require a paid license of more than $5M.
• All benchmarks and output numbers are sent via reproducible benchmarks/harnesses.

Interactive Dynamic Explainer