- Published on
A Guide to Open-Source Extraction Models in 2026
- Authors
- Name
- DeepHorus
Open-source models are now offering enterprise-grade capabilities to convert PDFs into structured JSON, unlocking vast pools of data for AI agents and large language models. This development addresses a critical bottleneck: most enterprise data remains trapped in PDFs, scans, and slide decks, inaccessible to automated systems. Proprietary APIs for document extraction can cost thousands per million pages and require sending data off-premise, making local open-source models an attractive alternative for cost and privacy reasons.
The phrase 'PDF to JSON' actually encompasses two distinct problems: schema-driven extraction and document parsing. Schema-driven extraction allows users to define specific fields, with the model returning values in valid JSON format—ideal for invoices, contracts, and forms. Document parsing, on the other hand, reconstructs the entire page layout, including reading order, tables, and formulas, outputting JSON or Markdown for use in retrieval-augmented generation (RAG) systems. Choosing the wrong category can waste significant time and resources.
Among the leading models, Datalab's lift stands out as a 9B vision-language model that accepts a JSON schema and returns matching JSON using schema-constrained decoding. On Datalab's benchmark of 225 documents, lift achieves 90.2% field accuracy with a median latency of 9.5 seconds, outperforming NuExtract3 (81.5%) and Qwen3.5-9B (76.3%). However, full-document accuracy—getting every field correct in one pass—remains challenging at only 20.9% for local models, compared to 95.9% for Datalab's hosted API.
NuMind's NuExtract 3, a 4B model based on Qwen, unifies structured extraction and content extraction (OCR to Markdown). It uses reinforcement learning to add extraction-specific reasoning and serves via vLLM with an OpenAI-compatible API. Meanwhile, IBM Docling, now hosted by the LF AI & Data Foundation, parses multiple document formats including PDF, DOCX, and images, outputting lossless JSON and Markdown while preserving layout and formulas. Docling runs locally for air-gapped environments.
The implications for enterprises are significant: open-source models reduce reliance on expensive cloud APIs and mitigate data privacy risks. Licenses vary, with lift's code under Apache-2.0 but its weights under a modified OpenRAIL-M license, free for research and startups under $5 million in funding or revenue. Commercial self-hosting requires a license, and competitive use against the Datalab API is prohibited. NuExtract 3 and Docling offer more permissive terms, encouraging broader adoption.
Reactions from the developer community have been positive, with many praising the ability to run extraction on local hardware. Datalab's lift, in particular, has been noted for its high field accuracy and schema-studio tool for testing. However, the low full-document accuracy underscores that this technology is still maturing, and careful validation is needed for production use cases.
Looking ahead, continued improvements in model architecture and training data are expected to push accuracy higher while reducing latency. Open-source document extraction is poised to become the standard approach for enterprises seeking to unlock their data for AI, as local models close the gap with proprietary solutions. The combination of cost savings, privacy, and community-driven innovation makes this a trend worth watching throughout 2026 and beyond.
