NuExtract3

Name: NuExtract3
Author: numind

by numind

Template-driven structured extraction from documents and images (VLM)

10Kdl/month

~4B (Qwen3.5-4B base)params

HuggingFace Run it on your own data, free

Identifiers

Model ID

numind/NuExtract3

Feature URI

mixpeek://document_extractor@v1/numind_nuextract3_v1

Overview

NuExtract 3 turns a vision-language model (built on Qwen3.5-4B) into a structured-extraction engine. Instead of returning free text, you hand it a schema, a JSON template describing the fields you want, and it fills that template from the input, whether the input is plain text or a document *image*. Because it sees the page, it reads scanned PDFs, invoices, forms, and screenshots directly, with no separate OCR stage, and it can also emit clean document-to-markdown.

On Mixpeek, NuExtract 3 is the extraction step that converts messy documents into queryable structured records: pull line items, dates, parties, and amounts into typed fields, then filter and search them like any other metadata.

Architecture

Vision-language model fine-tuned from Qwen3.5-4B (~4B params) for template-conditioned extraction. Takes an image or text plus a JSON schema/example and generates the populated structure, with light reasoning before committing values. Multilingual; handles document images at native resolution (inherited from the Qwen vision encoder), so layout and figures are read directly rather than flattened to a text dump.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so NuExtract3 runs
// on your side and the output is upserted through POST
// /v1/namespaces/{namespace_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection_id: "col_your_collection",
      documents: [
        {
          document_id: "asset-00412",
          // The model produces text, so it lands in payload. Give the
          // collection a text vector index and embed that text to make it
          // searchable rather than only filterable.
          payload: { extracted_text: modelOutput, source_key: "archive/2026/asset-00412" },
          vectors: { "text-embedding": embeddingOfModelOutput },
        },
      ],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// universal_extractor@v1 runs google/gemini-embedding-2
// (3072-d) over a bucket, with no inference of your own.

Capabilities

Schema/template-driven extraction: you define the fields, it fills them
Reads document images directly (scanned PDFs, invoices, forms): no separate OCR step
Document-to-markdown conversion for clean downstream chunking
Multilingual extraction with reasoning before value assignment
Open weights (Apache-2.0), self-hostable