Datalab Releases lift: A 9B Open-Weights Vision Model That Extracts Structured JSON From PDFs Using Schemas
June 23, 2026 – Datalab today announced the release of lift, a 9-billion-parameter open-weights vision model designed to extract structured JSON data from PDFs and images, guided by user-defined JSON schemas. In a 225-document benchmark, lift achieved a field accuracy of 90.2%, setting a new standard for schema-driven document parsing.
What is lift?
lift is a specialized vision-language model that takes a JSON schema as input and returns matching JSON output directly from PDFs or image files. Unlike general-purpose OCR or document understanding models, lift is built from the ground up for structured extraction, making it ideal for use cases like:
- Invoice and receipt parsing
- Form digitization and data entry
- Automated document processing pipelines
- Compliance and audit data extraction
By using an open-weights model, organizations can run lift on-premises or in air-gapped environments, avoiding the cost and privacy concerns of cloud-based API services.
How It Works
At its core, lift combines a vision encoder with a large language model decoder, trained on millions of document–schema–JSON triplets. The 9B-parameter architecture allows lift to understand complex layouts, tables, and handwritten text while respecting the structural constraints defined by the schema. The workflow is straightforward:
- Define a target JSON schema (e.g., fields for invoice number, date, total amount).
- Provide an image or PDF containing the document.
- Receive a structured JSON object conforming to the schema pattern.
- Field accuracy: 90.2%
- Schema compliance: >99% (outputs adhere to the schema structure)
- Average processing time: Under 5 seconds per page on a single A100-class GPU
- Hugging Face Model Hub:
datalab/lift-9b - GitHub Quickstart: github.com/datalab/lift
- API Sandbox: Available for testing on Datalab's website (limited to 10 documents per day)
Benchmark Performance
In Datalab's internal evaluation on a diverse set of 225 documents—including invoices, forms, and contracts—lift achieved:
These results represent a significant improvement over prior open-source document extraction models, which often require multiple passes or post-processing steps.
2026 Context: The Rise of Schema-Guided Extraction
In 2026, the demand for automated document processing has surged across industries—from healthcare (patient intake forms) to finance (loan applications) to logistics (shipping manifests). Traditional OCR and simple layout parsing are no longer sufficient; businesses need models that understand domain-specific structure and can output machine-readable data directly. lift joins a growing ecosystem of open-weights tools that prioritize accuracy, privacy, and control over proprietary cloud solutions.
Availability and Open Weights
lift's weights are available under an open license via Hugging Face and Datalab's official repository. The model supports PyTorch and ONNX runtime deployments. For now, lift is optimized for standard document formats; support for handwritten, multi-language, and rotated pages is planned for future releases.
"With lift, we're making it possible for any team—whether they're building a startup or running a Fortune 500 compliance department—to extract structured data from documents without shipping sensitive files to third-party APIs," said a Datalab spokesperson.
Getting Started
Developers can explore lift via:
Conclusion
As document parsing becomes a critical component of enterprise AI pipelines, lift offers a powerful, open-weights alternative to expensive, closed-source systems. With its schema-driven JSON extraction and 90.2% accuracy, lift is set to become a go-to tool for developers and data engineers in 2026.
This article was originally published on MarkTechPost and edited for clarity by AI.
via MarkTechPost