Datalab Lift: 9B parameter schema-first document extraction tool, compared with NuExtract3, LlamaExtract

Datalab released Lift, a document extraction tool centered on a schema-first approach. Unlike traditional workflows, Lift accepts a PDF or image along with a JSON Schema and directly outputs JSON that conforms to the schema, without first converting the document to Markdown and then having another model extract fields. It is based on a 9B parameter model and reads the rendered page image to generate final structured data. The article compares Lift with NuExtract3, LlamaExtract, Marker, and Docling, highlighting its simplified path and end-to-end processing capabilities. For developers or data engineers dealing with large volumes of documents (such as invoices, forms, contracts), Lift can reduce errors and latency from multi-stage extraction. By skipping intermediate conversions, output consistency and speed may be superior, especially for scenarios with strict JSON schema requirements. However, the article does not provide specific performance numbers or benchmark results; further real-world testing is needed to verify its claimed advantages.