Marker vs Docling: an independent comparison for RAG teams (November 2026)

Marker or Docling: the verdict

Marker vs Docling: comparison table

Compare all fields side by side, including certifications and hosting →

Marker vs Docling: what it costs per year

License: Docling is MIT, Marker splits code and model weights

Docling's code is under the MIT license, according to the README and the GitHub license file. The README adds that individual models carry their own licenses. On Hugging Face, the layout model docling-layout-heron and the GraniteDocling model are both tagged Apache-2.0 as of October 20, 2026. Docling was started by the AI for knowledge team at IBM Research Zurich and is hosted by the LF AI & Data Foundation.

Marker has two licenses. The code is Apache 2.0, and the README says it is free to use, including commercially. The model weights use a modified AI Pubs Open RAIL-M license. I read the license file itself. Its use restrictions bar use of the model, derivatives or the output in three commercial cases: when you or your employer had more than $5,000,000 in gross revenue in the prior year, when you have raised more than $5,000,000 in equity or debt in total, and when you offer a product or service that competes with something Datalab sells. The first two carry an exception for personal and research use. The competitor clause, as written, has no such exception.

Two details matter in practice. The license counts the entity you are affiliated with, so a developer at a large company is covered by the company's numbers. Its definition of distribution also includes offering the model as a hosted service through an API or web access, so a team that wants to sell a document API built on Marker should talk to Datalab first. This is a reading of the text, not legal advice.

What this means for you: if your company is small and does not sell document processing, Marker is usable without paying. Above the limits, or in a gray zone such as a document platform, plan for a license negotiation with Datalab. Docling has no such step.

Cost for 5 and 25 people over one year: free software, paid hosting

Neither project charges for software. The cost sits in compute, and in Marker's case in the hosted API. The numbers below are an assumption, not a measurement: I assume each person adds 2,000 pages a month to the pipeline, so 5 people send 10,000 pages a month (120,000 in a year) and 25 people send 50,000 pages a month (600,000 in a year). Your volume will differ. Prices are Datalab's published list prices as of October 20, 2026.

Datalab charges $4 per 1,000 pages for Convert in fast or balanced mode and $10 per 1,000 pages for accurate mode. Its Free plan includes $20 of monthly usage on a work email ($10 on a personal one) with 25 requests per minute, and no data processing agreement. Team costs $400 a month and includes $400 of usage, 400 requests per minute and a DPA. Enterprise is priced on request and adds single sign-on, your own VPC and air-gapped deployment. Processing in the EU adds 25 percent to usage. Opting into data retention gives a 25 percent discount, and qualifying startups get 33 percent in the first year. The two discounts do not combine.

For 5 people at 10,000 pages a month in balanced mode, list usage is $40 a month, or $480 a year. If the Free allowance is deducted from that usage, the bill is about $240 a year, but without a DPA. With a DPA you need Team at $400 a month, so $4,800 a year, of which most is unused allowance. For 25 people at 50,000 pages a month in balanced mode, usage is $200 a month and fits inside Team, so the year costs $4,800. In accurate mode, usage is $500 a month, and if usage above the included $400 bills at list rates, which I assume, the year costs $6,000. Adding the EU surcharge to accurate mode gives $625 a month, or $7,500 a year.

Running Marker or Docling yourself costs $0 in license fees below Marker's limits. Neither project publishes a server price, and it depends on your hardware, so I do not estimate it. Marker's README describes a GPU setup with Docker and the NVIDIA Container Toolkit, or a CPU and Apple silicon setup with llama.cpp. Our survey of document parsers lists per-page prices for Datalab, LlamaParse and Textract side by side.

What this means for you: at 5 people, self-hosting either tool is cheaper than any paid plan, and hosted Marker pays off only if you need the DPA or have no GPU. At 25 people, Team at $4,800 a year is small next to the engineering time of running a GPU service, so the choice is about control, not price.

Quality and speed: the benchmark that favors Marker is Marker's own

Marker's README reports results on olmocr-bench, a benchmark of 1,403 PDFs from Allen AI. Marker in balanced mode scores 76.0 overall and 83.5 on born-digital PDFs. In the same table MinerU's pipeline scores 72.7 and Docling scores 50.3 overall and 64.0 on born-digital files. Throughput on one B200 host is 2.9 pages per second for Marker balanced, 7.4 for fast mode, 23.7 for fast mode without OCR on CPU, and 2.1 for Docling. Datalab ran these tests, and it says Docling ran with its default pipeline.

An outside check exists for Marker only. The olmOCR 2 paper from Allen AI lists Marker 1.10.1 at 76.1 plus or minus 1.1 on olmOCR-Bench, dated September 2025, and MinerU 2.5.4 at 75.2 plus or minus 1.1. Docling is not in that table. So the Marker figure holds up independently, but the Docling figure of 50.3 comes only from a competitor's run. Docling offers other pipelines, including vision-model conversion with GraniteDocling, that the README table does not test.

Docling's own numbers are older. Its 2024 technical report measured the standard pipeline with OCR off on 225 pages. An Apple M3 Max with 4 threads processed 1.27 pages per second at a peak of 6.20 GB of memory, and an Intel Xeon E5-2690 processed 0.60 pages per second. The lighter pypdfium backend was faster, at 2.18 and 0.94 pages per second, but the report says table structure gets worse. Models have changed since, so these numbers may not match today's release.

Hardware, test sets and versions differ across these numbers, so I do not put them in one ranking. They support a narrower point. Marker has independent evidence for its overall score, and its CPU-only mode is fastest but gives up math and scans. Its README shows arXiv math at 0.0 with OCR off.

What this means for you: for mostly digital PDFs, both tools are plausible and the quality gap matters less than the README suggests. For scans, math or dense tables, test Docling's vision pipeline against Marker in balanced mode on your own pages first.

Formats, OCR and data handling: Docling reads more, and Marker's optional LLM step defaults to a cloud API

Marker converts PDF, images, PPTX, DOCX, XLSX, HTML and EPUB, and writes Markdown, JSON, a flat chunks format meant for RAG, or HTML. It extracts images, strips headers and footers by default, and runs OCR through a vision-language model on a local inference server that it starts itself, with vLLM on NVIDIA GPUs or llama.cpp elsewhere. Its README calls the built-in API server not robust and meant for small-scale use.

Docling's input list is longer: PDF, Word, Excel, PowerPoint, legacy Office formats through LibreOffice, OpenDocument, EPUB, Markdown, LaTeX, HTML, images, audio, video, email files and XBRL financial reports. It exports Markdown, HTML, DocTags, WebVTT and lossless JSON from one DoclingDocument model, supports several OCR engines such as EasyOCR, Tesseract and RapidOCR, and ships a command line tool, the docling-serve API server and an MCP server. LangChain, LlamaIndex, Crew AI and Haystack integrations exist.

Data handling differs most for teams in Germany, Austria and Switzerland. Docling's FAQ states that it uses no remote service and can run in isolated air-gapped environments once the model files sit in a local folder. Marker runs locally as well, but its optional LLM mode defaults to a Gemini model through Google's API, so page content leaves your network unless you point it at Ollama or another local endpoint. If you use Datalab's hosted API instead, a data processing agreement under Article 28 of the GDPR is needed, and the pricing page lists a DPA from the Team plan on. Datalab states zero data retention by default on its managed platform, and the pricing page offers EU data residency at 1.25 times usage.

What this means for you: for confidential documents, Docling with local models is the shortest path to a clean answer for a data protection officer. Marker is equally local only if you leave the LLM step off or run it on your own hardware.

What users complain about, from public GitHub issues

These points come from open issues I read on October 20, 2026. They show recurring pain, not how often it happens. Marker: a user reported running out of memory on a 246-page PDF with 8 GB of RAM even after sending 5 pages at a time (issue 825, opened August 6, 2025, updated September 7, 2026). Another reported a PDF of under 100 pages that hung for at least 7 hours (issue 888, September 17, 2025). A third asked about large files, where processing time grew with file size and blocked a worker (issue 487, January 15, 2025). A fourth reported stray newline characters inside Markdown math output (issue 755, June 13, 2025).

Docling: one issue reports that a conversion consumes all available memory and the process gets killed (issue 2779, December 13, 2025, updated May 1, 2026). Another reports memory growing by about 13 GB when one converter is reused on a 35-page PDF (issue 2209, September 5, 2025). Several users report that Markdown export marks every heading as level 2 and asks for a real heading hierarchy (issue 1023, February 19, 2025, and issue 287, November 9, 2024). A user converting a 300-page PDF asked how to speed it up (issue 1256, March 28, 2025). All of these were still open when I checked.

Both projects have open reports of memory trouble on large inputs. Heading levels are Docling's specific weakness in these threads, and long hangs are Marker's.

What this means for you: plan for batch limits, page-range splitting and a worker timeout whichever tool you pick, and put heading levels in your acceptance test if you chunk by section.

Switching checklist: re-parse everything, expect different chunk boundaries

Moving between the two means re-parsing your corpus, because the structures differ. Marker's JSON is a tree of typed blocks with a section hierarchy, and its chunks format is a flat list. Docling's lossless JSON is a DoclingDocument with its own chunkers, such as the HybridChunker. Chunk boundaries, table representation and image references will change, so the embeddings built on old chunks need rebuilding. Plan for a full re-index, not a patch.

Time depends on volume and hardware. As an illustration, 100,000 pages at the vendor's 2.9 pages per second for Marker balanced on one B200 host would take about 9.6 hours, and at 2.1 pages per second for Docling about 13.2 hours. At the Docling report's 1.27 pages per second on an M3 Max, it takes about 22 hours. This is arithmetic on published figures, not a measurement on your hardware.

What you lose depends on direction. Leaving Marker, you give up the chunks output, its block-level section hierarchy and the optional LLM correction prompt. Leaving Docling, you give up audio, video, email and most Office formats, its MCP server and its framework integrations. Run both on 50 representative pages, compare tables and headings by hand, and keep the old index until the new one passes.

Neither? LlamaParse for a hosted route, MinerU for another open-source one

If you do not want to run GPUs or accept Marker's license, LlamaParse from LlamaIndex is a hosted parsing API. Its pricing page lists Free with 10K credits, Starter at $50 a month with 40K credits, Pro at $500 a month with 400K credits, and states that 1,000 credits equal $1.25, with a Europe option. It suits teams that would rather pay per page than run servers.

If you want another self-hosted parser, MinerU from Shanghai AI Laboratory scores close to Marker in the Allen AI paper (75.2 against 76.1, each with a margin of 1.1). GitHub does not classify its license as a standard one, so read the license file before you commit. Neither Marker nor Docling comes with a service level on its own. Teams that need one should ask Datalab about Enterprise, which lists an MSA and SLA, or look at hosted vendors.

Who should pick Marker and who Docling

The bottom line

Start with Docling when licensing, offline operation or format coverage decide, and with Marker when you fall under its model license terms and want its output structure or Datalab's hosted API. In both cases, parse 50 of your own pages, check headings and tables by hand, and set batch limits before volume arrives.