Datalab Lift 9B Model Extracts Schema-Shaped JSON Directly

Datalab released Lift, a 9 billion parameter vision model designed for structured JSON extraction from PDFs and images. Lift operates on rendered page images and aims to directly output JSON that conforms to a user-provided schema in a single pass, differentiating itself from tools that first convert documents to intermediate formats like Markdown. This "schema-first" approach bypasses the common parse-then-extract workflow, where documents are first converted into representations like Markdown or HTML, and then a separate model extracts specific fields based on a schema.

The model's core functionality is to take a PDF or image along with a JSON schema and return schema-shaped JSON. This positions Lift as a specialized document extractor rather than a general OCR engine, a PDF-to-Markdown converter, or a comprehensive enterprise document review platform. Its output is described as "schema-shaped," meaning it directly provides the application-ready fields defined by the user's schema, such as invoice numbers, vendor names, or line items.

This direct extraction method contrasts with "parsing" tools, which create document-shaped intermediate representations. Examples of parsing tools include Docling, Marker, Unstructured, and Surya, which output formats like Markdown, HTML, or layout trees. Lift's innovation lies in collapsing the typical two-step process of parsing followed by extraction into a single visual extraction pass. This is intended to streamline workflows for applications that require structured data from visually complex documents.