What it is
Marker converts PDFs, images, Word, PowerPoint, Excel, HTML and EPUB files into Markdown, HTML, JSON or retrieval-ready chunks, keeping tables, equations, code blocks and reading order, and stripping running headers and footers. It is maintained by Datalab, which sells a hosted document API built on the same models.
Version 2.0.0, released on 20 July 2026, was a rewrite and a relicence: the code moved from GPL-3.0 to Apache 2.0. The model weights it downloads remain under a modified OpenRAIL-M licence, free for research, personal use and companies below $5M in revenue and funding; larger companies need a commercial licence from Datalab.
It runs in two modes. Balanced, the default on a GPU, uses Datalab's Surya vision model for layout and re-reads bad pages; fast, the default on CPU, uses a 20M-parameter layout model and the PDF's own text, calling the vision model only for equations and damaged blocks. An optional flag adds an outside LLM, Gemini by default, for multi-page tables and forms.
Our review
Marker is the open converter to reach for when PDFs, not Office files, are the problem. On olmOCR-bench, a third-party set of 1,403 PDFs, Datalab reports balanced mode at 76.0 overall and 83.5 on born-digital files, against 72.7 for MinerU's pipeline and 50.3 for Docling, while converting 2.9 pages a second on one B200 against MinerU's 0.54. These are the maintainer's own runs, but the harness is in the repository and the benchmark is not Datalab's, which is more than most rivals offer.
The licence needs reading twice. The July 2026 move from GPL-3.0 to Apache 2.0 removed the main objection to embedding the code in commercial software, but the code is useless without the Surya weights, and those are free only for companies under $5M in revenue and funding. Above that line the choice is a commercial licence or Datalab's API. Docling and MarkItDown carry no such condition.
The modes are a real trade-off rather than marketing tiers. Fast mode scores 66.6 at 7.4 pages a second but drops to 23.4 on arXiv maths, because it reads equations from the text layer; pure text-layer extraction runs on CPU at 23.7 pages a second and scores zero on maths and 14.3 on old scans. A corpus of clean contracts can run on CPU; a scanned archive or a library of papers needs a GPU and balanced mode.
Operationally, version 2 adds moving parts. Balanced mode spawns a vLLM inference server in Docker on NVIDIA GPUs, or llama.cpp elsewhere, and many thin workers share it. The bundled API server exposes only five parameters and is described in the README as not very robust, so a production service is either built around the Python API or bought from Datalab, whose own Chandra 2 model scores 85.8 on the same benchmark.
Where it earns its place
An ML engineer building search over 200,000 scientific papers
Turn PDFs with equations and tables into clean Markdown chunks
Run marker over the folder in balanced mode on one GPU host with --output_format chunks; equations come out as LaTeX. At the reported 2.9 pages a second on a B200, a 2 million-page corpus takes about eight days on one machine, or can be sharded across nodes with --num_chunks.
A legal-tech startup under $5M in revenue processing client contracts
Convert born-digital contracts to Markdown without GPUs or per-page fees
Use fast mode with --disable_ocr on ordinary CPU instances, which never starts the vision server. The weights are free at this company size, but the licence must be revisited once revenue or funding passes $5M.
A head of operations at a 2,000-person insurer comparing build and buy
Decide between self-hosting Marker and paying for Datalab's API
At 100,000 pages a month the API's Convert price of $4 per 1,000 pages is about $400, which the $400 Team plan covers and adds a BAA. Self-hosting needs a commercial weights licence at this size, plus GPU and on-call costs.
Before you commit
- Setup
- pip install marker-pdf (or marker-pdf[full] for Office and EPUB) on Python 3.10+ with PyTorch, then marker_single file.pdf. GPU mode also needs Docker and the NVIDIA Container Toolkit; CPU and Apple Silicon need llama.cpp's llama-server.
- Integrations
- Python API, CLI, a Streamlit app and a small FastAPI server; optional LLM boost through Gemini, Vertex, Claude, OpenAI-compatible endpoints, Azure OpenAI, OpenRouter or Ollama; a Modal deployment example.
- Data and privacy
- Runs locally, with no telemetry found in our reading of the repository; documents leave the machine only if --use_llm sends them to an outside model, Gemini by default. Weights under modified OpenRAIL-M. Datalab's hosted API has SOC 2 Type II on every plan and opt-in data retention.
- Cost at scale
- Self-hosted cost is GPU time: balanced mode reached 2.9 pages a second on one B200 using about 30% of the GPU, per the README. Datalab's API charges per 1,000 pages: $0.50 for lite Convert, $4 for fast or balanced, $10 for accurate, with 25% added for EU processing.
- Limits
- Nested tables and forms can fail, by the README's own account; the API server cannot use --use_llm; VRAM needs are not published. 385 open issues.
What's good
- 76.0 on olmOCR-bench in balanced mode, ahead of MinerU's pipeline and Docling
- Code relicensed to Apache 2.0 in July 2026
- CPU-only path for born-digital PDFs
- Markdown, HTML, JSON with bounding boxes, and chunk output
What's not
- Weights need a commercial licence above $5M revenue or funding
- Balanced mode needs Docker and a GPU-backed vLLM server
- Fast mode is weak on equations
- Bundled API server is basic and unhardened
Best for
RAG and data teams converting large volumes of complex or scanned PDFs on their own hardware.
Alternatives
Marker: common questions
- Is Marker worth it?
- Marker turns hard PDFs into usable Markdown better than any other open pipeline its makers measured, though the weights licence puts a revenue ceiling on that freedom. We rate it 4.3 out of 5.
- How much does Marker cost?
- Code free under Apache 2.0; weights free below $5M revenue and funding, commercial licence above. Hosted Datalab API: Convert $4 per 1,000 pages (fast or balanced); Team plan $400/month.
- What are the downsides of Marker?
- Weights need a commercial licence above $5M revenue or funding; balanced mode needs Docker and a GPU-backed vLLM server; fast mode is weak on equations; bundled API server is basic and unhardened.
- Who is Marker best for?
- RAG and data teams converting large volumes of complex or scanned PDFs on their own hardware.
- How much does Marker cost for a team or in production?
- Self-hosted cost is GPU time: balanced mode reached 2.9 pages a second on one B200 using about 30% of the GPU, per the README. Datalab's API charges per 1,000 pages: $0.50 for lite Convert, $4 for fast or balanced, $10 for accurate, with 25% added for EU processing.
- Is Marker safe for company data?
- Runs locally, with no telemetry found in our reading of the repository; documents leave the machine only if --use_llm sends them to an outside model, Gemini by default. Weights under modified OpenRAIL-M. Datalab's hosted API has SOC 2 Type II on every plan and opt-in data retention.
- What are the best Marker alternatives?
- The closest alternatives we have reviewed are Docling, MarkItDown and Unstructured.
Answers checked Oct 23, 2026.
Make Marker? Show this rating on your site
