Accurately extracting information from documents has been a decades-old dream. Many important workflows — from automated back-office processing to enterprise RAG — depend on it.

General-purpose LLMs promise to fulfill this dream, but they have drawbacks: they make mistakes, are expensive, and are difficult to use privately.

The solution: specialized SLMs.

In this talk, I will present NuExtract3 — a 4B open-source multimodal SLM specialized in document extraction. This reasoning model outperforms general-purpose LLMs with 70B+ parameters on both structured extraction (document to JSON following a schema) and OCR (document to Markdown).

I will explain how we built NuExtract3, demonstrate its capabilities, and discuss the future of information extraction and specialized SLMs.