LIFT: A Qwen3.5-Based VLM for PDF-to-JSON Extraction

Datalab has released LIFT, a vision-language model aimed squarely at one of the more tedious problems in document processing: pulling clean, structured data out of PDFs and other visually complex files. The model, versioned 1.0, is built on the Qwen3.5 foundation and outputs results directly as JSON.

The pitch is practical rather than flashy. Rather than positioning LIFT as a general-purpose multimodal assistant, Datalab frames it as a focused extraction tool — the kind of component you'd drop into a pipeline that ingests invoices, forms, reports, or scanned paperwork and needs reliable, machine-readable output on the other end.

Why it matters

Document extraction is a workhorse task across finance, legal, healthcare, and back-office automation, and it remains stubbornly hard because real-world documents rarely follow tidy layouts. A model that natively emits JSON lowers the integration friction for developers who would otherwise stitch together OCR, layout parsing, and post-processing logic by hand.

• Built on Qwen3.5, a recent and capable open foundation
• Outputs structured JSON suited to automated pipelines
• Released under the OpenRAIL license on Hugging Face

As an initial 1.0 release, LIFT arrives without published parameter counts or benchmark figures in its record, so teams will want to validate it against their own document sets before committing. Still, the arrival of a purpose-built, openly available extraction VLM is a useful addition for anyone tired of brittle, hand-rolled parsing stacks. The model is available now via its Hugging Face repository.