Vik Paruchuri Releases 9B Open Source Document Extraction Model With 90.2% Benchmark Score
A new open-source 9 billion-parameter model for extracting structured data from documents was released by Vik Paruchuri. On a 225-document benchmark he described, the model scored 90.2%, compared with 91.3% for Google's Gemini 3.5 Flash and 81.5% for document-extraction model NuExtract3, with median latency of 9.5 seconds.
Code, model weights and a Python package called lift-pdf were published alongside the release. The benchmark spans several document categories, with files ranging from 6 to 64 pages and including difficult extraction cases; Paruchuri also cited JSON schema support.
From the sources (4 posts)
@wightmanrRT @VikParuchuri: We're open sourcing a 9B model that extracts structured data from documents at near-frontier performance. - 90.2% on our…
@vikparuchuriWe're open sourcing a 9B model that extracts structured data from documents at near-frontier performance. - 90.2% on our bench, vs Gemini 3.5 Flash at 91.3% - Leads extraction models like NuExtract3 (81.5%) - 9.5s p50 timings - Pass JSON s
@vikparuchuriEverything you need to run it: pip install lift-pdf Model: Code: Blog:
@vikparuchuriDetails on the benchmark: 225 docs across several categories, from 6-64 pages. Includes common difficult extraction cases. More details here - .