In 2026, a guide to open-source extraction models will be available that convert PDF into JSON.

The majority of enterprise data is still contained in PDFs, slideshows and scans. The data cannot be used by large language models or agents until the JSON is structured. It is now standard to convert documents using open-source software on your own hardware.

Two different problems hide under the phrase ‘PDF to JSON.’ First, the first problem is schema-driven extractionYou define the fields and then a model will fill them up with values. Second, you can use a model to fill in the values. document parsingA model transforms a page to structured JSON, or Markdown. The majority of teams require one or both. The time you lose by choosing the wrong category is real.

Cost and privacy are important factors when it comes to open weights. Proprietary APIs cost up to thousands of dollars for every million pages and require that documents be sent outside the premises. Local models remove both constraints. Here are some models and tools worth evaluating. They’re grouped according to what they do.

Two phrases, two categories

The schema-driven extraction process takes your document, a JSON Schema and returns the values of all fields. It is useful for receipts, invoices and forms where the fields are known in advance.

Document parsing reconstructs a document. It exports JSON and Markdown after detecting layout, reading orders, tables, formulas, code, etc. You can use it to create clean corpora that you can then send off for RAG and agents.

Datalab lift

lift It is 9B model of the vision. DatalabThe team that created Marker and Surya. The JSON schema is passed to lift, which returns JSON that corresponds. Schema-constrained encoding assures that the output of lift is valid JSON. Model is built using Qwen and run either locally by Hugging Face through Qwen or remotely via vLLM.

The software handles multiple-page documents with a single operation, even if the values span several pages. It ships a CLI, a Python API, and a Streamlit ‘Schema Studio’ for building and testing schemas.

Pip Install Lift-PDF

Extract the schema after starting the vLLM Server
lift_vllm
lift_extract input.pdf ./output --schema schema.json

Lift import extract

Extract result ("document.pdf", "schema.json")
if the result.extraction value is None
    data = result.extraction  # dict matching your schema

Datalab benchmark 225 documents shows lift reaching 90.2% accurate field at median latency 9.5s. Field accuracy is higher than Qwen3.5-9B (76.3%) and NuExtract3 (81.5%). Gemini Flash 3.0 (91.3%) is ahead, followed by the Datalab API hosted on Datalab servers (95.9%). All local models have low accuracy for full documents, and lift is at 20,9%. It is still difficult to get every field correct in a single document.

Apache-2.0 is used for the code. Weights are licensed under a modified OpenRAIL M license. This is free for startups, research and personal use. A license is required for commercial self-hosting, and weights can’t be used in competition with Datalab API.

NuExtract 3 This is the 4B Vision-Language Model from NuMind. The model combines two separate tasks: content extraction and structured extraction. It requires an input, and you can provide a JSON-formatted template that describes which fields are required. This model has been trained using reinforcement learning, and you can turn on/off extraction-specific reasoning per request.

NuExtract 3 has a Qwen-based backbone and is multimodal. The vLLM API is OpenAI compatible, while a Python SDK can be downloaded via Pip Install Numind. NuMind considers it a model of open reference for content and structured extraction. You can check the model card Before commercial use, check the license conditions.

Categories 2 and 3: parsing JSON documents into structured Markdown

IBM Docling

Docling started at IBM Research and is now hosted by the LF AI & Data Foundation. This tool parses PDFs, DOCXs, PPTXs, XLSXs, HTML, images and many more. Markdown HTML lossless JSON and DocTags can be output. The DoclingDocument is its core, preserving layout, reading orders, tables and formulas in LaTeX.

For air-gapped systems, Docling is run locally. It is compatible with LangChain and LlamaIndex. Haystack and Crew AI are also supported. It is released under a permissive MIT License. IBM offers watsonx, a version managed by IBM.

IBM Granite-Docling-258M

Granite-Docling-258M IBM has a 258M model for vision-language. The Docling conversion pipelines perform a one-shot process of document conversion. It can handle OCR, table layouts, codes, equations and tables. With an A100 graphics card, the average page time is 0.35 second.

Model is based on Idefics3 Architecture, which includes SigLIP2 Encoders and Granite 165M Language Backbone. This software was released as Apache 2.0. IBM claims it was built to convert documents, and not for general image recognition.

OpenDataLab MinerU

MinerUThe, developed by OpenDataLab in conjunction with Shanghai AI Laboratory converts PDFs, images, DOCXs, PPTXs, and XLSXs to Markdowns and JSONs. It combines a processing-pipeline with a model of vision language. MinerU2.5Pro is a high-resolution model that parses complex layouts such as cross-page charts and tables.

MinerU changed its license recently. The license was changed from AGPL-3.0 (the previous version) to AGPL-3.0 (the new one). “MinerU Open Source License,” A custom license built on Apache 2.0, with conditions added. This change reduces the friction of commercial deployment.

Datalab Marker

Marker Datalab has a pipeline that converts documents to Markdown, JSON chunks and HTML. This tool supports HTML, EPUB, PDF, image PPTX and DOCX. Formats tables, forms and equations. It also formats links, code, inline math and other elements. An optional --use_llm Flag adds an improved language model for tables and forms.

Marker has a score of around 76.1 in the olmOCR Bench Suite, a third-party suite. The code of Marker is GPL-3.0 and the model weights are licensed under a modified AI Pubs OpenRAILM license. The weight license can be used for personal or research use and by startups with revenue under $2M. Datalab’s Managed Platform now uses a better OCR Model. Chandra. It is Apache-2.0.0 and can output HTML, Markdown or JSON.

Ai2 olmOCR 2

olmOCR 2 It is an OCR model with a vision-language specialization of 7B from the Allen Institute for AI. It can convert PDFs from a clean format into Markdown and maintain the order of reading. The program can handle tables, handwriting, equations and complex layouts with multiple columns. It is taught with reinforcement learning using verifiable reward signals, such as synthetic unit tests.

olmOCR2 scores an impressive 82.4 out of 100 on the olmOCR Bench. It is one of the best published results for this suite. Ai2 calculates that GPUs cost around $178 for every million pages. The toolkit is a combination of the allenai/olmOCR-2-7B-1025 Weights are Apache 2.0. The current English model has been developed.

DeepSeek DeepSeek OCR

DeepSeek-OCR DeepSeek will release an OCR open model in October 2025. It introduces “contexts optical compression,” It decodes the compact vision tokens into text. It can process longer documents using fewer tokens compared to typical vision-language models.

The decoder uses both a DeepEncoder and a 3B mix-of-experts, which activates approximately 570M variables per token. It can output plain text, HTML tables or structured JSON depending on what prompt is given. It is available under the MIT licence. DeepSeek OCR2 is a follow-up that will arrive in January 2026.

This is the general-purpose version: Qwen3VL

Qwen3-VL Alibaba’s model isn’t document-specific. This is an all-purpose multimodal model that’s used as the base for many other extraction models. The page can return Markdown code, JSON or both. Most sizes ship under Apache 2.0. This is an alternative when the specialized models are not suitable, but it requires more rapid engineering and has fewer output guarantee.

What are the choices?

Note: These benchmarks are from different suites, and therefore not comparable. Datalab’s schema extraction benchmark measures lift’s field accuracy at 90.2%. Content extraction is measured by the olmOCR Bench scores of olmOCR 2, (82.4), and Marker, (76.1). You should run your own files through each candidate prior to making a decision.