Nutrient

Nutrient (formerly PSPDFKit) is the deterministic document platform for agents. It combines reliable document processing infrastructure for agents, routing and governance, and interfaces for the human in the loop, so that agentic document workflows produce output the person signing off on the business outcome can rely on.

This organization is where we publish the small, document-specialized models behind that platform, together with the open benchmarks and self-service leaderboards we score them on. Same task, same data, same scorer, for our models and everyone else's.

At a glance

• Company: Nutrient, Raleigh, North Carolina. Formerly PSPDFKit. Website: nutrient.io
• Product entry point for extraction: Nutrient Data Extraction (parse PDFs, scans, images, and Office files; extract schema-defined fields with source citations and per-field confidence signals; free tier, no credit card)
• What we publish here: task-specific document models (open-weight and commercial), CC-BY-SA / CC-BY / MIT benchmark datasets, reproducible leaderboards, and no-signup demos
• Release pattern: an open-weight model (Apache-2.0, ONNX where possible) alongside a higher-accuracy commercial model, both scored on the same public benchmark. Commercial weights deploy on-prem, so documents never leave your environment
• Model-agnostic: everything here works next to any LLM, frontier or local, or none. Our platform overview for agents lives at nutrient.io/llms.txt

Why small document models

Most agents can pull data out of a document. Far fewer can prove where it came from, split a scanned batch back into its documents, tell an invoice from a remittance, or find the fields on a form at a quality level that stands up in an audit. Frontier LLMs are strong generalists and weak on exactly these document-shaped tasks: numbers inside tables, long page streams, dropout-ink forms, handwriting next to print.

We publish narrow models for those tasks and the benchmarks that show where each one wins and where it loses. The point is not to replace your model. It is to give an agentic workflow the deterministic pieces it needs to route, verify, and escalate to a human, and to let you check our numbers yourself.

What's here

Every leaderboard is scored by the same open scorer we use ourselves. Current numbers, including the cases where a cloud model beats us, live on the leaderboards, not in this README. To add a model, run the scorer and open a pull request on the leaderboard space.

Where these fit in an agentic document workflow

• Extract, then verify. Nutrient Data Extraction returns each field with a page reference, bounding box, source blocks, a confidence signal, and a match label. The grounding models score whether an extracted value is actually supported by the source, so fuzzy_match and not_found can be routed to human review while grounded values continue downstream. Confidence is a relative signal, not a calibrated probability, and we say so.
• Route and govern. Splitting and classification turn a mailbox, scanned batch, or upload folder into typed documents with known boundaries, the precondition for routing, governance, and audit trails.
• Keep a human in the loop. Form field detection and the rendered document viewers in our SDKs are the review surface: the human checks the real document, not a transcript of it. See AI Assistant and Enterprise Document Agents.

Use it in production

• Data Extraction: REST endpoint, getting-started guide, citations and confidence, published benchmarks and methodology. New accounts get 5,000 credits per month at no cost, no credit card, and signup is agent-compatible.
• Interactive extraction demos, no signup or key: nutrient-extraction-samples. Hover a field, see the exact source region.
• Build your own agent: MCP Server, SDKs, and Document Engine at nutrient.io/ai/infrastructure. Agent skills, including PDF-to-Markdown and the local document CLI: nutrient-skills.
• Commercial models on-prem or a custom build with our implementation teams: contact us.
• All open models and benchmarks in one place: nutrient.io/ai/models.

Citing and reusing

Datasets carry their own licenses (CC-BY-SA-4.0, CC-BY-4.0, or MIT, stated on each card). Open-weight models are Apache-2.0. If you use a benchmark or leaderboard result, link the leaderboard space so readers can see the scoring code and the full field at the time you cite it.

Contact

• Website: nutrient.io
• GitHub: PSPDFKit · PSPDFKit-labs
• X: @nutrientdocs
• Sales and implementation teams: nutrient.io/contact-sales