Datalab has launched Marker 2, a full rewrite of its open supply doc conversion pipeline. Marker converts PDF, picture, PPTX, DOCX, XLSX, HTML, and EPUB recordsdata into markdown, JSON, HTML, or chunks. The Datalab group rebuilt it round three parts shipped over the previous months: Surya OCR 2, a 20M-param quick structure mannequin, and a rebuilt pdftext that’s 3× sooner than the earlier one.
The principle outcome comes from olmOCR-bench, a third-party benchmark from Allen AI. Marker 2’s balanced mode scores 76.0% total and 83.5% on born-digital PDFs. It sustains 2.9 pages per second on a single B200 GPU. That’s over 5× the throughput of MinerU’s pipeline backend, which scores 72.7% at 0.54 pages per second. Docling scores 50.3% at 2.1 pages per second on the identical harness.
What’s New in Marker 2
Marker 2 exposes three conversion paths as an alternative of 1:
- balanced — the Surya VLM handles structure, and the entire web page is re-OCR’d every time embedded textual content is unhealthy. Highest high quality, greatest on GPU. 76.0% olmOCR-bench.
- quick — a light-weight rf-detr/onnx structure detector plus pdftext, with minimal, surgical VLM use. 66.6%, and much cheaper.
- –disable_ocr — pure text-layer extraction, no VLM calls in any respect. Runs totally on CPU. 43.6%, 23.7 pg/s.
Mode is now device-aware by default: balanced on GPU, quick on CPU/MPS, overridable with –mode. Full CPU help is the second structural change. quick –disable_ocr wants no GPU and no inference server, and the 20M structure mannequin nonetheless reads columns, tables and headers on CPU.
The third change is architectural, and it’s the one which produces the throughput numbers. Many skinny CPU employees share a single Surya inference server. The mum or dad course of budgets VLM concurrency throughout them, so throughput scales with server capability somewhat than per-process VRAM. Datalab stories that balanced mode sustains ~2.9 pg/s in opposition to a ~0.3 pg/s single-stream charge on the identical {hardware}.
Breaking modifications are price flagging earlier than an improve. Python 3.10+ is now required. Packaging moved from Poetry to uv, with hatchling because the construct backend, although pip set up marker-pdf is unchanged. The structured-extraction converter and extractors have been eliminated; Datalab factors customers to the hosted API or a –use_llm workflow as an alternative.
Comparability
The scoring benchmark is olmOCR-bench from Ai2: 1,403 PDFs with roughly 8,400 cross/fail unit assessments masking math rendering, desk construction, studying order, headers and footers, and outdated scans. The general rating is the macro-average throughout the 8 classes, computed with the official olmOCR-bench checker. Throughput is sustained concurrent pg/s on one B200 host, not single-stream latency.
A observe on provenance. olmOCR-bench is a third-party benchmark from Ai2, however each rating and throughput determine under comes from Datalab’s personal runs. All of them are reproducible by the open harness within the Marker repository, which ships competitor runners for MinerU, Docling and LiteParse alongside Marker’s personal.
These numbers additionally mirror one benchmark’s doc combine measured on a single {hardware} setup, so outcomes by yourself paperwork might differ. Groups evaluating these techniques ought to run the harness in opposition to their very own corpus, which is the one approach to know the way the 4 rank on the paperwork they really course of.
Marker 2 vs MinerU
MinerU’s pipeline backend is the closest architectural match. Each learn the PDF textual content layer and OCR selectively. On total rating, Marker balanced leads 76.0 to 72.7. On born-digital paperwork the 2 are successfully tied: 83.5 in opposition to 83.3.
The separation is throughput. Marker balanced sustains 2.9 pg/s in opposition to MinerU’s 0.54 pg/s, a 5.4× hole at a better rating. Marker quick sustains 7.4 pg/s, roughly 13.7× MinerU’s pipeline charge, however scores 6.1 factors under MinerU to do it.
MinerU additionally ships a VLM backend, which Datalab states scores increased than its pipeline backend. That backend is a full-page-VLM strategy and isn’t on this desk. AI groups evaluating MinerU ought to benchmark that path individually.
Marker 2 vs Docling
Docling is the widest margin among the many GPU pipelines. Marker balanced leads 76.0 to 50.3 total and 83.5 to 64.0 on born-digital, whereas additionally working sooner: 2.9 pg/s in opposition to 2.1 pg/s. Datalab notes Docling was run on its default pipeline, which makes use of the textual content layer for born-digital pages and OCR for picture areas.
Docling’s counterweight is governance and format breadth, not accuracy. The codebase is MIT-licensed, it originated at IBM Analysis, and it’s hosted as a mission within the LF AI & Information Basis. Its enter listing additionally extends previous paperwork into audio and electronic mail codecs.
Marker 2 vs LiteParse
LiteParse, from the LlamaIndex group, is a Rust doc parser. It doesn’t compete on the identical axis. On CPU it scores 22.4 total and 20.4 with OCR off, in opposition to Marker’s CPU-only 43.6.
However LiteParse with OCR disabled stories 1721 pg/s — roughly 73× Marker’s CPU mode, which is the tradeoff. Marker’s quick –disable_ocr runs a 20M structure mannequin on CPU and nonetheless recovers construction, which is why it greater than doubles a plain textual content dump’s rating. LiteParse has no structure mannequin and collapses on something non-linear.
Marker 2 vs the full-page VLM tier
The Datalab group emphasizes that Marker is designed as a pipeline somewhat than a VLM, clarifying that these are distinct instruments. On this analysis, their hosted Chandra 2 scores 85.8, whereas Gemini Flash 3.5 through API scores 76.4. Datalab’s Chandra repository additionally positions Ai2’s olmOCR 2 at 82.4 and dots.ocr 1.5 at 83.9 inside a separate desk. For scans, math-heavy pages, and attaining high accuracy, the VLM tier stays superior to all listed pipelines.
Marker’s balanced mode narrows the efficiency hole to simply 0.4 factors behind Gemini Flash 3.5 total, and it even outperforms it on born-digital paperwork by a margin of 83.5 to 79.1—with out requiring a per-page API name.
Per-category conduct
The mode you choose modifications the failure profile, not simply the rating. Every row is one olmOCR-bench class, scored throughout all three modes. Math is the sharp edge: quick mode reads equations from the PDF textual content layer as an alternative of VLM-OCRing them, so arXiv math falls from 83.9 to 23.4, and –disable_ocr scores 0.0 there by design. Outdoors the 2 math classes, outdated scans is the weakest cut up in each mode, topping out at 43.2.
Licensing
That is the place the 4 techniques diverge most for industrial groups:
- Marker: code is Apache 2.0. Mannequin weights use a modified AI Pubs OpenRAIL-M license — free for analysis, private use, and startups underneath $5M funding/income. Past that, industrial use of the weights requires a paid license.
- MinerU: now underneath the MinerU Open Supply License, primarily based on Apache 2.0 with added circumstances. A separate industrial license is required above 100M MAU or $20M month-to-month income, and on-line companies constructed on it should disclose that reality.
- Docling: MIT, with mannequin licenses tracked individually of their authentic packages.
- LiteParse: open supply, from run-llama, with LlamaParse positioned because the paid cloud path for onerous paperwork.
Use Case- Comparability
Rating alone doesn’t choose the software. Corpus sort, {hardware}, licensing band and output format resolve it. Attempt the interactive picker under to filter ten deployment eventualities by constraint and by software, and see which parser suits your use case.
Key Takeaways
- Marker 2 balanced scores 76.0% on olmOCR-bench at 2.9 pg/s — over 5× MinerU’s pipeline throughput.
- It beats Docling on each axes directly: 76.0% in opposition to 50.3%, and a pair of.9 pg/s in opposition to 2.1 pg/s.
- LiteParse trades construction for pace — 1721 pg/s with OCR off, however 20.4% in opposition to Marker’s 43.6% on CPU.
- Quick mode with –disable_ocr runs totally on CPU, no inference server, at 23.7 pg/s.
- Licensing splits the sector: Docling is MIT, MinerU stays free to $20M month-to-month income, and Marker’s weights want a paid license above $5M.
- All benchmark and throughput numbers ship with a reproducible benchmarks/ harness.
Interactive Dynamic Explainer
Hyperlinks: GitHub repo | Launch notes | Weblog publish | Announcement tweet
Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is dedicated to harnessing the potential of Synthetic Intelligence for social good. His most up-to-date endeavor is the launch of an Synthetic Intelligence Media Platform, Marktechpost, which stands out for its in-depth protection of machine studying and deep studying information that’s each technically sound and simply comprehensible by a large viewers. The platform boasts of over 2 million month-to-month views, illustrating its reputation amongst audiences.
