NuExtract3
NuMind's 4B Apache 2.0 vision-language model that turns documents into JSON by template or into Markdown, with switchable reasoning.
In short. NuExtract3 reads a scanned page, form or invoice and either fills in a list of fields you define, such as invoice number and total, or rewrites the whole page as clean Markdown. It is small enough for one consumer GPU, has a mode that thinks before it answers, and is free under Apache 2.0.
Summary
NuExtract3 is NuMind's 4.54B-parameter vision-language model for document understanding, released on 19 May 2026 and fine-tuned from Qwen3.5-4B. It does two jobs with one model. For structured extraction you give it text or page images, a JSON template whose values name the field types, and optional instructions or examples, and it returns JSON with null for anything it cannot find. NuMind says the template language has 20 field types, including verbatim strings, dates, countries and currencies. For content extraction, it turns page images into Markdown, with HTML tables and LaTeX for math. It was trained with supervised fine-tuning and then reinforcement learning, and reasoning can be switched on or off.
On NuMind's internal structured-extraction benchmark of about 600 documents, scored by leaf accuracy, the card reports an average of 0.651 against 0.538 for Gemma-4-E4B-it, 0.479 for Qwen3.5-9B, 0.417 for Qwen3.5-4B and 0.387 for Nemotron-3-Nano-Omni, with 27 unparseable outputs against 31 to 229 for those four. NuMind says it plans to release that benchmark. For Markdown it uses two tests it designed itself, a language-model judge on challenging layouts and a check of how well a separate model can extract fields from the Markdown, and in both NuExtract3 is ahead of similarly sized models. NuMind notes that reasoning hurt several rival models, which fell into repetition loops, and that its own judge-based test has flaws. All of these are developer-reported.
The weights are 9.32 GB in BF16, and NuMind also publishes FP8, W4A16 and GGUF builds, the smallest being a 2.78 GB Q4_K_M file that needs a separate 0.68 GB vision projector. The card serves the model with vLLM, with custom code enabled, up to 99 images per prompt and a 131,072-token window, and optional multi-token-prediction speculative decoding. Multi-page PDFs must be rendered to page images first. The card recommends starting in non-reasoning mode and turning thinking on only for hard documents. In the 30 days to 11 November 2026, Hugging Face counted about 40,700 downloads for the BF16 repository and 21,700 for the GGUF one.
Good for
- Turning invoices, receipts, forms and contracts into JSON that follows a schema you define, on a single GPU
- Preparing scans and PDFs as Markdown for a retrieval-augmented generation pipeline
- Hard layouts such as split or overlapping tables, where the optional reasoning mode can help
Watch out for
- The benchmarks are NuMind's own, the structured-extraction set is not yet public, and the Markdown tests rely on a language-model judge that NuMind says has flaws
- Thinking mode costs time and tokens, averaging about 2,000 thinking tokens per document in the card's benchmark, so the card suggests starting with it off
- PDFs have to be turned into page images by you, vLLM needs custom-code and image-limit flags, and a GGUF run also needs the separate vision projector file
- It is specialized for extraction rather than general chat, and the card gives no minimum memory
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the NuExtract3-Q4_K_M (GGUF) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| NuExtract3 (BF16) BF16 | 4.54B | 10 GB | 5.22 GB | 3.13 GB | Laptop or phone |
| NuExtract3-FP8 FP8 | 4.54B | 10 GB | 5.22 GB | 3.13 GB | Laptop or phone |
| NuExtract3-W4A16 INT4 | 4.54B | 10 GB | 5.22 GB | 6.07 GB | Laptop or phone |
| NuExtract3-Q4_K_M (GGUF) INT4 | 4.54B | 10 GB | 5.22 GB | 3.06 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is NuExtract3?
- NuExtract3 reads a scanned page, form or invoice and either fills in a list of fields you define, such as invoice number and total, or rewrites the whole page as clean Markdown. It is small enough for one consumer GPU, has a mode that thinks before it answers, and is free under Apache 2.0.
- Can I use NuExtract3 commercially?
- Yes. NuExtract3 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is NuExtract3 open source?
- NuExtract3 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run NuExtract3 on my own computer?
- Yes. The smallest variant, NuExtract3-Q4_K_M (GGUF), needs about 3.06 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was NuExtract3 released?
- NuExtract3 was released by NuMind on May 19, 2026.
- How long a context window does NuExtract3 have?
- NuExtract3 supports a context window of up to 128K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- NuExtract3 model card (Hugging Face) huggingface.co
- NuExtract3: The Reasoning Open-Source OCR & Structured Extraction LLM (NuMind blog) numind.ai
- NuExtract repository (GitHub) github.com
- NuExtract3-GGUF model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Chandra OCR 2 Datalab | 5.3B | 256K | OpenRAIL-M (Datalab) (Conditional) | 3.66 GB Laptop or phone | |
| LightOnOCR-3 LightOn | 853M – 4.54B | 256K | Apache 2.0 (Permissive) | 0.589 GB Laptop or phone | |
| GLM-OCR Z.ai (Zhipu AI) | 900M | – | MIT (Permissive) | 0.621 GB Laptop or phone | |
| PaddleOCR-VL Baidu | 960M | – | Apache 2.0 (Permissive) | 0.662 GB Laptop or phone |