# NuExtract3 by NuMind: Open-Weight Model Source: https://savrn.com/models/nuextract3 Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- ## Runs On What it takes to serve NuExtract3 (4.5B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them. | Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits | | --- | --- | --- | --- | --- | --- | | 16-bit | 9.1 GB | 10.9 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 8-bit | 4.5 GB | 5.4 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 4-bit | 2.3 GB | 2.7 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the [SAVRN Index](https://savrn.com/ai-index/pricing/gpus), read Oct 25, 2026. [NuExtract3 on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/nuextract3/gpus) ## SAVRN's Notes on NuExtract3 NuExtract3 pulls structured data out of documents: give it a page and a JSON template, and it fills the fields. It also converts pages to markdown, so one 4.5B-parameter model covers both OCR and extraction. NuMind built it on a Qwen 3.5 vision base with a 262,144-token context. It needs about 11 GB at 16-bit, one mid-range GPU. Apache 2.0 allows commercial use. For invoice, contract and form processing, template-driven extraction like this is easier to audit than free-form generation; start with your ten messiest documents. SAVRN Research, 2026-09-18 ## Model Card By NuMind, published under apache-2.0, revision c99dc8f5641b. [API / Platform](https://nuextract.ai/) | [Blog](https://numind.ai/blog) | [Discord](https://discord.gg/3tsEtJNCDe) | [GitHub](https://github.com/numindai/nuextract) NuExtract3 is a unified 4B vision-language reasoning model for document understanding. It combines strong structured information extraction with high-quality image-to-Markdown conversion, making it suitable for extraction pipelines, OCR, and RAG preprocessing for all types of documents such as scans, receipts, forms, invoices, contracts or tables. Try it out in [the space!](https://huggingface.co/spaces/numind/NuExtract-3-4B) ### Overview - Structured extraction: input (text/images) + JSON template + instructions --> JSON output - Markdown conversion: input (text/images) --> Markdown - Multimodal inputs: text, images, or text + images. - Multilingual documents. - Reasoning and non-reasoning inference modes. - Template generation for structured extraction from natural language or input document. ### Benchmark results ### Structured Extraction [Read the full model card (2,355 words)](https://savrn.com/models/nuextract3/card) ## Configuration Architecture Qwen3_5ForConditionalGeneration Context length (tokens) 262,144 Layers 32 Hidden size 2,560 Feed-forward size 9,216 Attention heads 16 Key/value heads 4 Head dimension 256 Vocabulary size 248,320 Model type qwen3_5 ## Identity and Version Repository numind/NuExtract3 Publisher NuMind Task Image to text Modality Image and text Library transformers Parameters 4.5B parameters Languages vlm, ocr, rag Revision c99dc8f5641b866aa0192b6ea78f84bf9f3535f1 First published 2026-04-29 Last updated 2026-08-20 ## Files and Weights 29 files, 9.3 GB in total. The weights are 2 files totalling 9.3 GB in safetensors. Weights2 files · 9.3 GB Configuration6 files · 62.6 KB Tokenizer2 files · 20.0 MB Documentation3 files · 40.9 KB Other15 files · 3.6 MB Repository1 file · 2.2 KB Every file | File | Type | Size | SHA-256 | | --- | --- | --- | --- | | model.safetensors | Weights | 9.1 GB | aca0a9d61da5 | | model_mtp.safetensors | Weights | 241.2 MB | 7f993d7b896c | | .eval_results/extractbench.yaml | Configuration | 1.6 KB | — | | .eval_results/olmocrbench.yaml | Configuration | 502 B | — | | config.json | Configuration | 2.9 KB | — | | generation_config.json | Configuration | 115 B | — | | model.safetensors.index.json | Configuration | 56.3 KB | — | | processor_config.json | Configuration | 1.2 KB | — | | LICENSE | Documentation | 11.3 KB | — | | README.md | Documentation | 25.5 KB | — | | TYPES.md | Documentation | 4.1 KB | — | | chat_template.jinja | Other | 6.8 KB | — | | extraction_performance_no_thinking.png | Other | 257.4 KB | 873703147b4a | | extraction_performance_thinking.png | Other | 293.5 KB | cc50cef0718e | | header.svg | Other | 110.4 KB | — | | logo.png | Other | 1.1 MB | eacf669413cc | | logo_nuextract.svg | Other | 8.9 KB | — | | markdown.png | Other | 1.1 MB | b898b030d41e | | markdown_base.png | Other | 166.2 KB | 0475336139bd | | markdown_reasoning.png | Other | 161.5 KB | 733f46ad4c82 | | md2st.svg | Other | 88.6 KB | — | | ocr_benchmark_specialized.png | Other | 227.6 KB | 7c6e21084c03 | | ocr_preferences.svg | Other | 71.2 KB | — | | st.svg | Other | 39.7 KB | — | | task_instructions_markdown.txt | Other | 2.7 KB | — | | task_instructions_structured.txt | Other | 11.0 KB | — | | .gitattributes | Repository | 2.2 KB | — | | tokenizer.json | Tokenizer | 20.0 MB | 87a7830d63fc | | tokenizer_config.json | Tokenizer | 1.2 KB | — | ## License and Download License apache-2.0 Access Open weights, no gate Download size 9.3 GB [Download from NuMind](https://huggingface.co/numind/NuExtract3) Released by NuMind through its official repository on Hugging Face. [Read the license](https://www.apache.org/licenses/LICENSE-2.0). ## Built From - Derived from [Qwen/Qwen3.5-4B](https://savrn.com/models/qwen3-5-4b) ## Evaluations Each result is shown as reported, with the conditions its reporter stated. None is a SAVRN measurement. A comparison lines two results up only when their configuration, unit and setup are all stated and identical. | Benchmark | Conditions | Result | Reported by | Revision | Date | | --- | --- | --- | --- | --- | --- | | allenai/olmOCR-bench | Task old_scansMetric old_scansSetup old_scans.jsonl sub-score (present/absent/order); markdown mode, non-thinking, greedy, 170 DPI. NuExtract3 leads the field on present (41.6) — see source for the full sub-score breakdown.Comparison conditions not established | 37.8 | [ocr-bench — old_scans multi-model comparison](https://github.com/davanstrien/ocr-bench/blob/99f7550c/experiments/olmocr-bench-oldscans/BENCHMARKING.md) Reported by a third party | Evaluated revision not stated | 2026-06-27 | | llamaindex/ExtractBench | Task longMetric longSetup Pipeline name: nuextract3_extract; served checkpoint numind/NuExtract3 on vLLM, one-shot template-guided extraction (the JSON Schema is converted to a NuExtract template)Comparison conditions not established | 37.72 | [ExtractBench](https://huggingface.co/datasets/llamaindex/ExtractBench) Reported by a third party | Evaluated revision not stated | 2026-08-20 | | llamaindex/ExtractBench | Task meanMetric meanSetup Pipeline name: nuextract3_extract; served checkpoint numind/NuExtract3 on vLLM, one-shot template-guided extraction (the JSON Schema is converted to a NuExtract template)Comparison conditions not established | 82.35 | [ExtractBench](https://huggingface.co/datasets/llamaindex/ExtractBench) Reported by a third party | Evaluated revision not stated | 2026-08-20 | | llamaindex/ExtractBench | Task mediumMetric mediumSetup Pipeline name: nuextract3_extract; served checkpoint numind/NuExtract3 on vLLM, one-shot template-guided extraction (the JSON Schema is converted to a NuExtract template)Comparison conditions not established | 76.76 | [ExtractBench](https://huggingface.co/datasets/llamaindex/ExtractBench) Reported by a third party | Evaluated revision not stated | 2026-08-20 | | llamaindex/ExtractBench | Task shortMetric shortSetup Pipeline name: nuextract3_extract; served checkpoint numind/NuExtract3 on vLLM, one-shot template-guided extraction (the JSON Schema is converted to a NuExtract template)Comparison conditions not established | 88.06 | [ExtractBench](https://huggingface.co/datasets/llamaindex/ExtractBench) Reported by a third party | Evaluated revision not stated | 2026-08-20 | ## Memory Requirements | Precision | Weights in memory | | --- | --- | | As published | 9.3 GB | | 16-bit | 9.1 GB | | 8-bit | 4.5 GB | | 4-bit | 2.3 GB | Weights only, from the published parameter count; the key-value cache and runtime add to this. ## Questions About NuExtract3 ### How much GPU memory does NuExtract3 need? About 10.9 GB at 16-bit and 2.7 GB at 4-bit: the weights (4.5B parameters) plus a working margin. A long context needs more. ### What is the cheapest GPU to run NuExtract3 on? At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists. ### Can I use NuExtract3 commercially? Yes. NuExtract3 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors. ### What is NuExtract3's context length? 262,144 tokens, from the maximum position embeddings in its published configuration. ## Similar Models Model · Image to text ### [Qwen3-VL-8B-Instruct-NVFP4](https://savrn.com/models/qwen3-vl-8b-instruct-nvfp4) [LEE, HYEOKSEUNG](https://savrn.com/model-publishers/jeildlwlrma) This is an NVFP4 quantized version of Qwen3-VL-8B-Instruct, a powerful vision-language model for multimodal understanding and generation tasks. The following modules were excluded from quantization to maintain model quality: - lmhead (language model head) - Visual encoder modules (model.visual.) - MLP gate projections (.mlp.gate$) For faster inference, you can use this model with vLLM: This quantized model maintains high quality for vision-language tasks while significantly reducing memory usage. The SmoothQuant technique helps preserve model accuracy during quantization. Typical quality degradation is 2-5% compared to the full-precision model. 1. Calibration: Used 512 samples from the… Open weights apache-2.0 5.7B parameters 262,144 tokens [View model](https://savrn.com/models/qwen3-vl-8b-instruct-nvfp4) Model · Image to text ### [granite-vision-3.3-2b](https://savrn.com/models/granite-vision-3-3-2b) [IBM Granite](https://savrn.com/model-publishers/ibm-granite) Granite-vision-3.3-2b is a compact and efficient vision-language model, specifically designed for visual document understanding, enabling automated content extraction from tables, charts, infographics, plots, diagrams, and more. Granite-vision-3.3-2b introduces several novel experimental features such as image segmentation, doctags generation, and multi-page support (see Experimental Capabilities for more details) and offers enhanced safety when compared to earlier Granite vision models. The model was trained on a meticulously curated instruction-following data, comprising diverse public and synthetic datasets tailored to support a wide range of document understanding and general image… Open weights apache-2.0 3B parameters 131,072 tokens [View model](https://savrn.com/models/granite-vision-3-3-2b) Model · Image to text ### [kosmos-2-patch14-224](https://savrn.com/models/kosmos-2-patch14-224) [Microsoft](https://savrn.com/model-publishers/microsoft) This Hub repository contains a HuggingFace's transformers implementation of the original Kosmos-2 model from Microsoft. Use the code below to get started with the model. This model is capable of performing different tasks through changing the prompts. First, let's define a function to run a prompt. Here are the tasks Kosmos-2 could perform: Once you have the entities, you can use the following helper function to draw their bounding bboxes on the image Open weights mit 1.7B parameters 2,048 tokens transformers [View model](https://savrn.com/models/kosmos-2-patch14-224) Model · Image to text ### [GLM-OCR](https://savrn.com/models/glm-ocr) [Z.ai](https://savrn.com/model-publishers/zai-org) Join our WeChat and Discord community Use GLM-OCR's API GLM-OCR is a multimodal OCR model for complex document understanding, built on the GLM-V encoder–decoder architecture. It introduces Multi-Token Prediction (MTP) loss and stable full-task reinforcement learning to improve training efficiency, recognition accuracy, and generalization. The model integrates the CogViT visual encoder pre-trained on large-scale image–text data, a lightweight cross-modal connector with efficient token downsampling, and a GLM-0.5B language decoder. Combined with a two-stage pipeline of layout analysis and parallel recognition based on PP-DocLayout-V3, GLM-OCR delivers robust and high-quality OCR performance… Open weights mit 1.3B parameters 131,072 tokens transformers [View model](https://savrn.com/models/glm-ocr) Model · Image to text ### [GLM-OCR](https://savrn.com/models/glm-ocr-2) [Unsloth AI](https://savrn.com/model-publishers/unsloth) Join our WeChat and Discord community Use GLM-OCR's API GLM-OCR is a multimodal OCR model for complex document understanding, built on the GLM-V encoder–decoder architecture. It introduces Multi-Token Prediction (MTP) loss and stable full-task reinforcement learning to improve training efficiency, recognition accuracy, and generalization. The model integrates the CogViT visual encoder pre-trained on large-scale image–text data, a lightweight cross-modal connector with efficient token downsampling, and a GLM-0.5B language decoder. Combined with a two-stage pipeline of layout analysis and parallel recognition based on PP-DocLayout-V3, GLM-OCR delivers robust and high-quality OCR performance… Open weights mit 1.3B parameters 131,072 tokens transformers [View model](https://savrn.com/models/glm-ocr-2) Model · Image to text ### [GLM-OCR](https://savrn.com/models/glm-ocr-3) [Manny](https://savrn.com/model-publishers/manny-the-one) Join our WeChat and Discord community Use GLM-OCR's API GLM-OCR is a multimodal OCR model for complex document understanding, built on the GLM-V encoder–decoder architecture. It introduces Multi-Token Prediction (MTP) loss and stable full-task reinforcement learning to improve training efficiency, recognition accuracy, and generalization. The model integrates the CogViT visual encoder pre-trained on large-scale image–text data, a lightweight cross-modal connector with efficient token downsampling, and a GLM-0.5B language decoder. Combined with a two-stage pipeline of layout analysis and parallel recognition based on PP-DocLayout-V3, GLM-OCR delivers robust and high-quality OCR performance… Open weights mit 1.3B parameters 131,072 tokens transformers [View model](https://savrn.com/models/glm-ocr-3) ## NuMind [All models and datasets](https://savrn.com/model-publishers/numind) ## Versions - [c99dc8f5641b](https://savrn.com/models/nuextract3/versions/c99dc8f5641b) · current 2026-09-18 ## Explore More - [All image to text models](https://savrn.com/models/tasks/image-to-text) - [All models under apache-2.0](https://savrn.com/models/licenses/apache-2-0) - [Model comparisons](https://savrn.com/models/comparisons) - [The model directory](https://savrn.com/models) - [Open model prices by host](https://savrn.com/ai-index/pricing/open-models) ## Source - Repository metadata, read 2026-10-20. - [Hugging Face record](https://huggingface.co/numind/NuExtract3) - [How the hub is built](https://savrn.com/model-hub/methodology)