# NuMind

> NuMind is a Paris-based, Y Combinator-backed AI company building multimodal large language models that turn messy documents - PDFs, scans, spreadsheets, handwritten forms - into clean, structured JSON. Its NuExtract platform and open-source models let developers and enterprises extract exactly the fields they define, in any language, either as a hosted API or deployed privately on their own infrastructure.

- **Founded:** 2022
- **Headquarters:** Paris, France
- **Founders:** Etienne Bernard (Co-Founder & CEO), Samuel Bernard (Co-Founder & CTO)
- **Team size:** ~13 employees
- **Products:** NuExtract Platform, NuExtract 2.0 (VLM family), NuExtract (foundation model), NuNER / NuNER Zero, NuMarkdown
- **Notable:** Raised $3M seed in 2023 led by Flybridge as a YC S22 company, Released NuExtract, an open-source foundation model for structured text-to-JSON extraction, Released NuExtract 2.0, a multimodal, multilingual vision-language extraction model family (2B/4B/8B)

## Products & services

- **NuExtract Platform** — A hosted API and SaaS platform that extracts structured JSON from documents - text, PDFs, spreadsheets, scans - in any language, powered by NuMind's in-house extraction models. Available as SaaS or enterprise/on-prem deployment.
- **NuExtract 2.0 (VLM family)** — A family of open-source vision-language models (2B, 4B, 8B) built on the Qwen-VL family, specialized in multimodal, multilingual structured extraction. Reads layouts, tables, and images, not just text.
- **NuExtract (foundation model)** — The original open-source text-to-JSON foundation model, fine-tuned from phi-3-mini on a synthetic dataset designed to mitigate hallucinations. Takes a text passage plus a JSON template.
- **NuNER / NuNER Zero** — Open-source named-entity-recognition backbone models, including a zero-shot variant, for token classification and entity extraction.
- **NuMarkdown** — Open-source model/tooling for converting documents to clean Markdown as part of the extraction pipeline.

## Achievements

- Raised $3M seed in 2023 led by Flybridge as a YC S22 company
- Released NuExtract, an open-source foundation model for structured text-to-JSON extraction
- Released NuExtract 2.0, a multimodal, multilingual vision-language extraction model family (2B/4B/8B)
- Improved a client's field accuracy from 30% to 95% on complex Japanese health-claim documents
- Models are generally SOTA in their category and released under MIT / Apache 2.0 licenses

## Latest updates

- **2025-07** — NuMind launches NuExtract 2.0, a vision-language model for extracting structured data from PDFs and images.
- **2025** — NuExtract platform advances toward a v3 generation, powered by the NuExtract 2.0 PRO model.
- **2025** — Open-source NuExtract 2.0 model collection (2B, 4B, 8B, GGUF) published on Hugging Face.

## Links

- Website: https://numind.ai
- LinkedIn: https://www.linkedin.com/company/numind-ai
- Twitter/X: https://x.com/numind_ai
- GitHub: https://github.com/numindai

---

Profile page: https://yespress.io/numind-yc-s22
Published by YesPress — https://yespress.io
Last updated: 2026-07-30