Marker Review 2026

High-performance open-source tool that converts PDFs, images, PPTX, DOCX, XLSX, HTML, EPUB, and other documents to markdown, JSON, chunks, or HTML with deep-learning-powered OCR, layout detection, and optional LLM cleanup.

Marker is a free open-source document conversion pipeline for permitted research, personal, and qualifying startup use, with Datalab managed API pricing at $4 per 1,000 pages for Fast/Balanced mode, $6 per 1,000 pages for High Accuracy and extraction workflows, and custom self-hosting pricing. Marker is designed to turn complex documents into clean, structured outputs for AI, search, analytics, and knowledge-base workflows. Its core job is converting PDFs and other document formats into markdown, JSON, chunks, or HTML while preserving useful document structure such as headings, tables, forms, equations, inline math, links, references, images, and code blocks. Five concrete implementation facts define the tool: it is installable as the Python package marker-pdf; the README requires Python 3.10+; non-PDF support is enabled through the fuller marker-pdf[full] installation; the project supports local execution through CLI commands such as marker_single and marker; and it can also be used through Python APIs, a Streamlit GUI, or a lightweight FastAPI server. The input coverage is broad for a document AI converter: PDF, image, PPTX, DOCX, XLSX, HTML, and EPUB files are all described as supported inputs, while markdown, HTML, JSON, and chunks are documented outputs. The chunk output is especially relevant for retrieval workflows because it flattens top-level blocks for easier ingestion into downstream RAG or search systems. Marker also includes specialized modes for narrower tasks, including table-only conversion, OCR-only conversion, and beta structured extraction. Its optional --use_llm mode can connect to services such as Gemini, Google Vertex, Ollama, Claude, OpenAI-compatible endpoints, and Azure OpenAI to improve hard cases like table merging across pages, inline math, table formatting, and form value extraction. Local deployment is practical for developers who can manage PyTorch and model dependencies, but resource planning matters: the README states Marker may use about 5GB of VRAM per worker at peak and about 3.5GB on average. Licensing also matters. The repository states that the code is GPL-3.0 and that model weights are free for research, personal use, and startups under $2M in funding or revenue, while broader commercial licensing or removing GPL requirements requires Datalab commercial licensing. For teams that do not want to operate the stack directly, Datalab offers a managed API, high-volume batch arrangements, and commercial self-hosted or on-premise options with custom terms.