Marker 2 Scores 76.0 on olmOCR-bench at 5x MinerU’s Throughput

Datalab has released Marker 2, a complete rewrite of its open-source document conversion tool. The new version scores 76.0% on olmOCR-bench (a third-party benchmark from Allen AI) and sustains 2.9 pages per second on a single B200 GPU. That puts it ahead of MinerU (72.7%, 0.54 pg/s), Docling (50.3%, 2.1 pg/s), and LiteParse (22.4% on CPU) across both accuracy and throughput in Datalab's own published runs.

What happened

Datalab rewrote Marker from the ground up and shipped Marker 2 on July 21, 2026. Marker converts PDFs, images, PPTX, DOCX, XLSX, HTML, and EPUB files into markdown, JSON, HTML, or chunked output. The rewrite is built on three components the team shipped in preceding months: Surya OCR 2, a 20-million-parameter layout model, and a rebuilt pdftext library that runs 3x faster than the previous version.

Fact	Detail
Release date	July 21, 2026
Benchmark	olmOCR-bench (Allen AI / Ai2), 1,403 PDFs, ~8,400 tests
Marker 2 balanced score	76.0% overall, 83.5% born-digital
Marker 2 balanced throughput	2.9 pg/s on one B200 GPU
MinerU pipeline score / throughput	72.7% / 0.54 pg/s
Docling score / throughput	50.3% / 2.1 pg/s
LiteParse CPU score	22.4% overall, 1,721 pg/s (OCR off)
Python requirement	3.10+

Three conversion modes, not one

Marker 2 ships with three distinct operating modes. The right choice depends on your hardware and how much accuracy you need.

• balanced: Surya VLM handles layout and re-OCRs any page with poor embedded text. Scores 76.0% on olmOCR-bench. Best on GPU.
• fast: Uses a lightweight rf-detr/ONNX layout detector plus pdftext, with minimal VLM calls. Scores 66.6% and sustains 7.4 pg/s. Much cheaper to run.
• –disable_ocr: Pure text-layer extraction. No VLM calls, no GPU required. Scores 43.6% but hits 23.7 pg/s on CPU.

The mode now defaults automatically based on device: balanced on GPU, fast on CPU or Apple MPS. You can override it with the --mode flag.

How the throughput gains were achieved

The architecture change is the key engineering story here. Instead of each worker process running its own model, many thin CPU workers now share a single Surya inference server. The parent process controls VLM concurrency across all workers, so throughput scales with server capacity rather than per-process VRAM.

The result: balanced mode sustains roughly 2.9 pg/s concurrent, compared to about 0.3 pg/s in single-stream mode on the same hardware. That multiplier is the gap you see in the benchmark table.

How does Marker 2 compare to MinerU, Docling, and LiteParse?

Tool / Mode	olmOCR-bench overall	Born-digital score	Throughput (pg/s)
Marker 2 balanced	76.0%	83.5%	2.9
Marker 2 fast	66.6%	N/A	7.4
Marker 2 –disable_ocr	43.6%	N/A	23.7
MinerU pipeline	72.7%	83.3%	0.54
Docling default	50.3%	64.0%	2.1
LiteParse (CPU, OCR off)	22.4%	N/A	1,721

A few nuances worth noting. MinerU also ships a VLM backend that Datalab says scores higher than its pipeline backend, but those numbers are not in this table. Teams serious about MinerU should benchmark that path separately. Docling’s advantage is governance breadth: it is MIT-licensed, originated at IBM Research, and sits inside the LF AI and Data Foundation. Its format support also extends to audio and email. LiteParse (from LlamaIndex, written in Rust) hits 1,721 pg/s with OCR off but collapses on non-linear layouts since it has no layout model.

For reference on the full-page VLM tier: Datalab’s own hosted Chandra 2 scores 85.8, Gemini Flash 3.5 via API scores 76.4, olmOCR 2 sits at 82.4, and dots.ocr 1.5 at 83.9. Marker 2 balanced is competitive with Flash 3.5 while running as a local pipeline rather than an API call.

Breaking changes before you upgrade

1. Confirm your environment is on Python 3.10 or higher before upgrading.
2. Note the packaging change from Poetry to uv with hatchling as build backend. The pip install marker-pdf command itself is unchanged.
3. Remove any code relying on the structured-extraction converter or extractors. Datalab has dropped those; the hosted API or a --use_llm workflow replaces them.
4. Run Datalab’s open benchmark harness against your own document corpus before committing. The olmOCR-bench mix is 1,403 PDFs from Ai2’s collection, which may not reflect your content type.

Why it matters

PDF-to-structured-text pipelines are infrastructure for a huge range of AI integration projects: RAG systems, document search, data extraction from contracts or invoices. Accuracy and throughput are usually in tension, and Marker 2’s balanced mode closing the gap against full-page VLM systems while running locally is a meaningful practical shift.

The throughput story also matters for cost. At 5x MinerU’s rate on the same GPU, a team processing millions of pages per month can either cut compute spend or handle the same load on fewer machines. For teams already tracking document parsing options in our AI news coverage, this is one of the more concrete benchmark drops we have seen this year.

Our take

Be careful with the provenance here. Every number in this comparison comes from Datalab’s own runs, not a neutral third party. The harness is open and reproducible, which is the right move, but “we beat the competition” benchmarks should always be treated as a starting point rather than settled truth. Docling in particular is almost certainly better than 50.3% in real deployments with its VLM mode enabled, and MinerU’s VLM backend was explicitly excluded.

That said, the architectural change is real and the throughput gains are plausible. If you are running a document parsing pipeline at scale and have been waiting for a local option that can keep pace with GPU capacity, Marker 2’s concurrent-worker design is worth testing against your own files. Just run the harness on your own corpus before making any infrastructure decisions.

Source: Marktechpost