Datalab has released lift, a 9B open-weights vision model built for extracting structured data from PDFs and images. The model is designed to produce JSON that matches a provided schema rather than free-form text.
The release highlights schema-constrained decoding for valid structure and trained abstention, allowing the model to return null when a field is absent instead of inventing an answer. Datalab reports 90.2% field accuracy on a 225-document benchmark.
Document extraction remains a major bottleneck for enterprise AI workflows, so an open model focused on reliable structure could be useful for teams handling forms, reports, and scanned records.