Docling vs Marker vs MinerU for document parsing

Choose the input path and output contract first

Docling converts supported files into DoclingDocument, with a lossless JSON serialization and exports such as Markdown and HTML. Its format guide also lists JSONL chunks for RAG. Legacy DOC/XLS/PPT and RTF require LibreOffice; audio transcription requires the asr extra, and video additionally requires ffmpeg. Check dependencies for the actual formats in your collection.

Marker supports a JSON block tree as well as Markdown, HTML and chunks. MinerU 4 uses a shared document model, but its entrypoints expose different exports: mineru-kit parse accepts markdown, middle_json and zip, while the self-hosted V1 API additionally exposes structured_content. Its Python renderer supports further formats. A mineru parse --json response contains command status, content and locators; it is not the intermediate document JSON. Do not build an ingestion contract around an interchangeable meaning of JSON.

Review code and weights separately

Docling code is MIT licensed, and its README directs users to each model’s original licence. Granite-Docling-258M is a documented VLM option whose official model card lists Apache-2.0. This does not establish the licence of every OCR engine, enrichment model or replacement VLM.

Marker code is Apache-2.0. Its README summarizes the weights as free for research, personal use and startups under $5 million in funding/revenue, but MODEL_LICENSE is more specific: Attachment A restricts use when the user, employer or affiliated entity had more than USD 5 million gross revenue in the prior year or has raised more than USD 5 million total equity or debt funding, with personal-use/research exceptions for those two clauses. A separate clause restricts use by entities offering a product or service competing with the licensor or its affiliates. Sections 7–8 impose attribution and share-alike terms, including terms addressing output. Commercial and broader licences may be available from Datalab; the code licence alone does not settle model permissions.

MinerU’s current LICENSE.md is Apache 2.0 with additional terms. A separate commercial licence is required before continuing commercial use if you and affiliates, consolidated, exceed either 100 million monthly active users or USD 20 million total monthly revenue. Online services for third parties must prominently acknowledge MinerU in their interface or public documentation. Failure to obtain a required licence or provide that attribution automatically terminates the granted rights. The official Torch model bundle and MinerU2.5-Pro-2605-1.2B VLM card label their weights Apache-2.0; the ONNX bundle says original model licences and attribution remain applicable. Check the complete deployed model stack rather than treating the repository licence as blanket clearance for weights.

Match the inference backend to the machine

Docling documents CPU-only installation and GPU configuration. Its installation guide warns that newer PyTorch wheels are unavailable for Intel Macs and gives a compatible installation path requiring Python 3.12 or lower. Its model catalog lists device support per stage; TableFormer currently disables MPS, so an Apple GPU path for one stage does not imply that all stages use it.

Marker 2 defaults to balanced mode on GPU and fast mode on CPU/MPS. OCR uses a local inference server that starts on demand: vLLM through Docker on NVIDIA, or llama.cpp on CPU/Apple Silicon. Disabling OCR avoids the VLM server but also skips equations and scanned pages. Budget for the backend dependencies when evaluating a local installation.

MinerU 4.1 raises the minimum Python version to 3.11. Basic can run without a VLM; Standard and Advanced also need VLM inference. The current ONNX path documents automatic acceleration on compatible NVIDIA, AMD and Intel GPUs on Windows/Linux, and MPS/Metal on macOS. Check the exact GPU, driver, inference backend and parsing tier before deploying. These documentation paths do not establish throughput or memory requirements for your files.

Configure extraction for scans, tables and math

Docling offers OCR engines including EasyOCR, Tesseract, RapidOCR and macOS OCR, with engine-specific dependencies. Its catalog documents TableFormer for cell structure. Formula enrichment is an additional step that extracts LaTeX and can be enabled with --enrich-formula.

Marker reconstructs tables from the PDF text layer where possible, uses OCR for scanned tables and can fall back to the VLM. Inline math is converted to LaTeX automatically in balanced mode; fast mode needs force_ocr or ocr_inline_math for that behavior. Optional --use_llm processing can merge tables across pages.

MinerU’s Basic tier covers OCR, formulas and tables, while Standard combines small models with a VLM. Flash with --ocr-mode txt reads the native PDF text layer without a fallback for scans; OCR mode requires the relevant models. A successful text-only conversion therefore does not establish that scanned input will be handled by the same configuration.

Check the API you will actually call

Docling documents official LangChain and LlamaIndex extensions. The LlamaIndex integration includes a reader and a node parser for downstream chunks. docling-serve exposes conversion over REST with synchronous results or asynchronous tasks.

Marker’s small-scale FastAPI server does not expose --use_llm or --disable_ocr; use the CLI or Python API for those paths. MinerU 4.1 uses asynchronous /v1/tasks on its main API and Router. Synchronous /v1/file_parse is served by a separate file-parse-server wrapper that connects to the API server, rather than by the main API itself. Its Python SDK, local renderer and hosted online service have distinct interfaces and output contracts.

Separate hosted prices from self-hosted licences

IBM prices its managed Docling API in Resource Units. One RU covers 1,000 pages or objects, or 50 million characters for text and spreadsheet inputs. The current Pay Per Use details start at USD 4.36/RU, while the same page’s headline advertises USD 4 per 1,000 pages. Confirm the applicable order price before budgeting. A free trial is available, and Docling’s managed-service guide documents the hosted REST API.

Datalab lists Free pay-as-you-go usage, Team at $400/month with $400 included usage, and custom Enterprise pricing. Free monthly allowance is $20 for a work email or $10 for a personal email. Convert fast/balanced starts at $4 per 1,000 pages and accurate at $10; EU processing adds 25%, and spreadsheets use cell-based billing. Datalab warns that some balanced/accurate jobs exceed the included agent-compute allowance and cost more. These are hosted processor prices. Its platform runs a version of Chandra, so hosted results do not establish local Marker behavior or model-licence pricing.

MinerU provides an online interface and hosted API. Its official API documentation specifies a 200 MB/200-page file limit and 1,000 pages of highest-priority processing per account per day; later pages receive lower priority. No public monetary rate was verified on the reviewed pages, so no hosted price or unlimited-free claim is made here.

Package metadata checked November 7, 2026 lists docling 2.148.0, released November 5; marker-pdf 2.1.0, released October 29; and mineru 4.1.7, released November 4. This comparison uses current official documentation, including MinerU 4.1 service changes. Pin packages, models and matching backend documentation before deployment.

How to choose

Consider Docling when the unified document representation and documented LangChain/LlamaIndex path fit your pipeline. Consider Marker when its conversion outputs and configurable OCR/LLM workflow fit, after reviewing the model restrictions and server limits. Consider MinerU when tiered local parsing, the document library and its service tools match the application, after reviewing the additional licence terms and export boundaries. These are SOTA’s documentation-based judgments, with no hands-on benchmark or accuracy ranking. Evaluate representative scans, tables and formulas using the exact version, models and hardware you plan to deploy.

Sources & method

Official documentation was checked on 2026-11-07. Selection advice is our assessment. Interfaces and requirements may change; check the linked source for the version you intend to use.

• Docling README: Python, CLI, code licence and separate model terms
• Docling MIT licence
• Docling input formats, dependencies and output formats
• Docling installation, CPU-only path, platforms and Intel Mac caveat
• Docling model catalog: OCR, table structure and per-stage device support
• Docling formula enrichment and LaTeX extraction
• Official Granite-Docling-258M model card and Apache-2.0 licence
• Docling official LangChain integration
• Docling official LlamaIndex reader and node parser
• Docling REST service and synchronous/asynchronous conversion
• Docling managed-service guide
• Docling for IBM watsonx hosted service, trial and published price
• Marker README: formats, OCR, tables, math, Python API and server limits
• Marker Apache-2.0 code licence
• Marker model licence: Attachment A restrictions and sections 7–8
• Datalab hosted plans, allowances, processor rates and commercial model licensing
• MinerU 4 README: input formats, tools and local native parsing
• MinerU current licence: Apache 2.0 plus thresholds and service attribution
• MinerU 4 installation, Python range and platform-specific dependencies
• MinerU parsing tiers, CPU/GPU engines and resource planning caveats
• MinerU model sources and separate remote service types
• Official MinerU Torch model bundle: Apache-2.0 label
• Official MinerU2.5-Pro-2605-1.2B VLM card: Apache-2.0 label
• Official MinerU ONNX model bundle: original licences and attribution remain applicable
• MinerU local Python parser and remote API SDK
• MinerU output contract: entrypoint exports, renderer formats and JSON distinctions
• MinerU official online interface
• MinerU hosted API documentation: file limits and daily processing priority
• docling 2.148.0 on PyPI
• marker-pdf 2.1.0 on PyPI
• mineru 4.1.7 on PyPI
• MinerU current CLI and separate API service tools

How SOTA reviews projects · Suggest a correction · Download the evaluation checklist