AI for documents: Datalab Lift vs. the competition

Lift, from Datalab, is a 9-billion-parameter model that extracts structured data from PDFs and document images. The company has published a comparison pitting its tool against four popular alternatives: NuExtract3, LlamaExtract, Marker, and Docling.

How Lift works

Lift is built schema-first: you pass in a PDF or a page image along with a JSON Schema, and the model returns ready-to-use structured JSON with no intermediate steps.

Most competing pipelines work differently. First, a converter turns the document into Markdown using OCR plus layout recognition. Then a separate language model pulls the required fields out of that text. Lift collapses both stages into one: it analyzes rendered page images directly and outputs the result in the required format.

Key specs:

• Model size: 9B parameters
• Input formats: PDFs and page images
• Output format: JSON that strictly follows the supplied JSON Schema
• Architecture: vision-based, with no intermediate Markdown
• Competitors in the comparison: NuExtract3, LlamaExtract, Marker, Docling

How the competitors differ

Marker and Docling are document converters. They specialize in faithfully reproducing page structure as Markdown or HTML. That helps when you need a text layer for search, indexing, or downstream processing by a language model, but on its own it doesn't give you structured data.

NuExtract3 and LlamaExtract are closer to Lift in purpose: both take a schema and return structured JSON. However, they typically operate on text that has already been converted, not on the raw visual representation of the page.

Lift's approach is a niche one. The model gives up versatility (there's no "just read the document" mode) in exchange for accuracy and directness in scenarios where the data schema is known in advance.

When the schema-first approach wins

A schema-first architecture makes the most sense in production pipelines where the output structure is defined up front: extracting fields from invoices, contracts, medical records, customs declarations, and bank statements.

In these cases, a two-step "convert to Markdown → LLM extraction" pipeline adds an unnecessary link: a parser error in the first step degrades quality in the second. Lift eliminates that risk by working directly with the visual representation of the page and targeting the final format from the start.

The limitation is a hard dependency on the schema. Lift isn't suited to unstructured tasks or cases where you need "all the text in the document." For those scenarios, Marker or Docling remain the better choice.

What it means

Datalab's comparison draws a new dividing line in the document AI market: general-purpose converters versus specialized extractors. For teams with clear output requirements, schema-first tools can significantly simplify the pipeline and reduce the number of moving parts in production systems.

How does the Lift AI model extract data?

Lift is built schema-first: you pass in a PDF or a page image along with a JSON Schema, and the model returns ready-to-use structured JSON with no intermediate steps.

What are the alternatives to Datalab Lift?

The main competitors are NuExtract3, LlamaExtract, Marker, and Docling. The difference is that they first convert the document to Markdown via OCR and then extract the data.