Last updated: 2026-08-25
NuMind released NuExtract3 in April 2026, a 4-billion-parameter open-weight vision-language model that outperforms a larger 9-billion-parameter rival model by 17 points on NuMind's roughly 600-document internal benchmark. Founded in 2022 by Etienne and Samuel Bernard, the Boston company converts PDFs, scans and forms into structured JSON via schema-guided extraction.
About NuMind
NuMind is a Boston-based document-AI company founded in 2022 by cousins Etienne Bernard (CEO, former head of machine learning at Wolfram Research) and Samuel Bernard (CTO, former CTO of Make.org). It launched through Y Combinator's Summer 2022 batch as a tool for building custom NLP models and has since narrowed entirely to structured document extraction, raising a total of $3.5M in seed funding led by Flybridge. The core product, the NuExtract Platform, converts unstructured documents such as PDFs, scans, spreadsheets and free text into structured JSON by pairing a document with a user-defined schema template, and leaves a field blank rather than guessing when information is missing from the source. In April 2026 NuMind open-sourced NuExtract3, a vision-language model fine-tuned from Alibaba's Qwen3.5-4B that can run on a single consumer GPU. It doubles as a document-to-Markdown converter for retrieval-augmented-generation preprocessing pipelines, and the open weights let teams self-host with no API keys, no per-call fees and no data leaving their own infrastructure. Pricing is usage-based rather than tiered: template creation and testing in the web playground are free, and the hosted API bills per token. Enterprises needing tighter data control can deploy the platform privately on their own cloud or a dedicated hosted instance, with additional privacy safeguards described in the product's data policy. The product suits data scientists, ML engineers and backend developers who need programmatic, schema-guided extraction rather than a point-and-click OCR tool for occasional use. It competes most directly with Reducto, Mindee and Veryfi in the document-parsing API space, and with LlamaIndex's ingestion layer for teams building RAG pipelines.
Pricing
Usage-based, not tiered. Web UI (template creation, teaching examples, playground testing) is free. API extraction is billed per token: $1/M for input, $5/M for output, roughly $0.001 for a single text page up to $0.15 for 100 scanned A4 pages. Enterprise private/on-premises deployment uses tiered volume pricing with optional fine-tuning, available through sales. The open-weight NuExtract3 model can be self-hosted for zero ongoing cost.
Key Features
- Schema-guided JSON extraction: The NuExtract Platform API pairs any document (PDF, scan, spreadsheet, free text) with a user-defined JSON template, and NuMind reports it beats GPT-4.1 by over 9 F-score points on its internal benchmark, running on the NuExtract 2.0 PRO model.
- Open-weight NuExtract3 model: A 4-billion-parameter vision-language model, less than half the size of the 9-billion-parameter rival it beats by a 17-point margin, measured across roughly 600 diverse documents in NuMind's own test set.
- Document-to-Markdown conversion: NuExtract3 converts scanned pages, forms and tables directly into clean Markdown alongside its structured JSON extraction mode, built for retrieval pipeline preprocessing.
- Teaching examples and reasoning modes: A project-based web playground lets users add in-context teaching examples to steer extraction accuracy without fine-tuning, with optional reasoning and non-reasoning inference paths.
- Private and on-premises deployment: Enterprises can choose private-cloud or on-premises deployment instead of the shared SaaS platform; production documents are deleted within two weeks of processing and the models are not trained on submitted data.
- Python SDK for platform integration: An MIT-licensed nuextract-platform-sdk on GitHub lets developers wire the hosted API into extraction pipelines directly from Python instead of hand-rolling HTTP calls.
Pros
- NuExtract3 ships as fully open weights under an Apache 2.0 license, letting teams self-host a document-extraction model instead of depending on a hosted API for every call.
- The platform returns a blank field instead of a guessed answer when information is missing from a document, reducing silent extraction errors compared to generic LLM prompting.
- Usage-based API pricing means small extraction jobs cost a small fraction of a cent per page instead of a flat monthly seat fee.
- The founding team has a research background in ML and NLP products (a former Wolfram Research ML lead and a former Make.org CTO), which shows in NuExtract3 beating far larger open-weight rivals on NuMind's own benchmark.
Cons
- The hosted NuExtract Platform API caps context at roughly 32,000 tokens (about 60 text pages or 20 image pages), forcing long contracts or multi-document batches to be chunked before extraction.
- No public G2, Trustpilot or Capterra review presence as of August 2026, so buyers have little independent, non-vendor evidence of reliability at scale.
- NuMind runs on a lean team, which YC's own profile lists at six employees, a much thinner base than venture-scale rivals such as Reducto for enterprise support commitments.
Frequently Asked Questions
How much do you pay for NuMind?
NuMind's NuExtract Platform is priced by usage rather than a flat subscription: extraction API calls cost $1 per million input tokens and $5 per million output tokens, which works out to roughly a tenth of a cent for a single text page and around $0.15 for 100 scanned pages. Template creation, testing and the playground are free in the web UI. Enterprise customers can move to private or on-premises deployment with tiered volume pricing and optional fine-tuning, available through sales.
Is NuMind free to use?
The web interface, including template creation, teaching examples and playground testing, costs nothing. Actual document extraction through the API is billed per token rather than offered as a fixed free quota, so there is no monthly cap to hit, only usage cost once documents are sent. Teams that want zero ongoing cost can instead self-host the open-weight NuExtract3 model under its own license.
What should you use instead of NuMind?
Reducto and Mindee both run managed document-parsing APIs with broader table and layout handling than NuMind's schema-first approach, worth a look if the documents are less structured. Veryfi specializes in receipt and invoice OCR with built-in accounting integrations that NuMind does not offer out of the box. LlamaIndex fits teams that want document ingestion built into a wider RAG framework rather than a standalone extraction API.
Is NuMind better than Reducto?
NuMind wins on cost control and self-hosting: an openly licensed model a team can run entirely on its own infrastructure, plus usage-based API pricing instead of a flat subscription. Reducto positions itself as a broader agentic document platform with wider out-of-the-box layout and table parsing, suiting teams that do not want to hand-define an extraction schema. Pick NuMind when self-hosting or per-token cost control matters more than zero-schema setup.
How long does it take to get going with NuMind?
Sign up on the NuExtract Platform, create a project, and define a JSON template describing the fields to extract, a process most users finish in a single sitting. Adding a handful of teaching examples and testing in the playground before hitting the API typically takes under an hour. Teams that want to skip the hosted API entirely can instead pull the open-weight NuExtract3 model from Hugging Face and run it locally.
Top Alternatives
- Reducto: Pick NuMind if per-token cost control or self-hosting matters more than zero-schema, out-of-the-box layout parsing.
- Mindee: Pick Mindee if built-in invoice and receipt templates matter more than defining a custom JSON schema by hand.
- Veryfi: Pick Veryfi if accounting-system integrations for receipts and invoices matter more than general-purpose document extraction.
- LlamaIndex: Pick LlamaIndex if document ingestion needs to sit inside a wider retrieval pipeline builder instead of a dedicated extraction API.