--- license: apache-2.0 base_model: ATH-MaaS/OvisOCR2 language: - multilingual library_name: mlx pipeline_tag: image-text-to-text tags: - ocr - document-parsing - markdown - tables - formulas - multimodal - vision-language - qwen3.5 - mlx - apple-silicon - oq - oqe6 - quantized --- # OvisOCR2-oQe6 > **Apple Silicon Optimized oQe6 MLX Quantized Release** This repository contains an **oQe6 mixed-precision MLX quantized** version of **OvisOCR2**, optimized for fast and memory-efficient document understanding on Apple Silicon. The original **OvisOCR2** model was developed by **ATH-MaaS**. This repository provides an optimized **MLX/oQe6 conversion only** and **does not include any additional training or fine-tuning**. Using **oQe6 sensitivity-aware mixed-precision quantization**, this release preserves the excellent OCR and document parsing capabilities of the original model while significantly reducing memory usage and improving inference efficiency on Apple Silicon. OvisOCR2 is a compact 0.8B end-to-end document parser that converts document images directly into structured Markdown and achieves state-of-the-art results on OmniDocBench v1.6 and PureDocBench. 0 --- # About OvisOCR2 OvisOCR2 is a compact **Vision Language Model (VLM)** specialized for **end-to-end document parsing**. Unlike traditional OCR pipelines that separately detect layouts, recognize text, and reconstruct documents, OvisOCR2 directly converts a document page into structured Markdown while preserving natural reading order. The model is capable of extracting: - 📄 Plain text - 📐 Mathematical formulas (LaTeX) - 📊 Tables (HTML) - 🖼 Images and visual regions - 📚 Complex layouts - 📰 Multi-column documents - 📑 Scientific papers - 📋 Forms - 📖 Books - 📃 Receipts - 📜 Scanned documents The model is based on **Qwen3.5-0.8B** and is trained using a combination of: - Real-world document data - Synthetic document generation - Supervised Fine-Tuning (SFT) - Reinforcement Learning (RL) - On-Policy Distillation (OPD) OvisOCR2 achieves an **overall score of 96.58 on OmniDocBench v1.6**, becoming the first end-to-end model to top that benchmark, and also leads PureDocBench with an Avg3 score of 75.06. 1 --- # Quantization This release uses **oQe6 mixed-precision quantization**. ## Specifications - **Format:** MLX - **Quantization:** oQe6 - **Method:** Sensitivity-Aware Mixed Precision - **Target Platform:** Apple Silicon - **Inference Engine:** MLX / oMLX Unlike traditional fixed-bit quantization, **oQe6 automatically assigns precision according to layer sensitivity**, preserving critical components while compressing less sensitive regions. Benefits include: - Better OCR accuracy retention - Lower memory usage - Faster inference - Higher throughput - Excellent Apple Silicon optimization --- # Recommended Settings For the best OCR quality: ```yaml temp: 0.0 top_p: 1.0 top_k: 1 max_tokens: 4096 ``` For difficult or noisy scans: ```yaml temp: 0.1 top_p: 0.95 top_k: 20 max_tokens: 4096 ``` OCR is generally deterministic, so greedy or near-greedy decoding is recommended for maximum transcription accuracy. --- # Example Usage ```python from mlx_vlm import load, generate from PIL import Image model, processor = load("yugeshkarunamurthy/OvisOCR2-oQe6") image = Image.open("document.png") prompt = """ Extract all readable content from the document. Output Markdown preserving reading order. Render tables as HTML. Render formulas using LaTeX. """ response = generate( model=model, processor=processor, image=image, prompt=prompt, temperature=0.0, ) print(response) ``` --- # Optimized For This release is optimized for: - Apple M1 - Apple M2 - Apple M3 - Apple M4 Compatible with: - MLX - MLX-VLM - oMLX - Local OCR Applications - Document Processing Pipelines --- # Model Highlights - End-to-End OCR - Page-level Document Parsing - Markdown Generation - HTML Table Extraction - LaTeX Formula Recognition - Scientific Paper Parsing - Forms and Receipts - Complex Multi-column Documents - Reading Order Preservation - High OCR Accuracy --- # Intended Use OvisOCR2-oQe6 is well suited for: - OCR - PDF Digitization - Scientific Paper Extraction - Book Digitization - Invoice Processing - Receipt Processing - Form Parsing - Research Automation - Knowledge Base Construction - RAG Preprocessing - Markdown Conversion - Digital Archives --- # Hardware Recommendations Recommended systems: - Apple M1 Pro / Max / Ultra - Apple M2 Pro / Max / Ultra - Apple M3 Series - Apple M4 Series The compact 0.8B model is lightweight and runs comfortably on most Apple Silicon devices while benefiting from additional memory for larger documents and higher throughput. --- # About oQe6 Quantization oQe6 is a sensitivity-aware mixed-precision quantization technique designed to preserve model quality while significantly reducing memory requirements. Rather than assigning identical precision to every layer, oQe6 analyzes the sensitivity of individual modules and allocates higher precision only where it has the greatest impact. Benefits include: - Better OCR accuracy retention - Improved document layout understanding - Lower RAM usage - Faster inference - Excellent Apple Silicon performance --- # Original Model The original **OvisOCR2** is an end-to-end document parsing model built upon **Qwen3.5-0.8B**. Notable features include: - State-of-the-art document OCR - Markdown document generation - HTML table extraction - LaTeX formula recognition - Natural reading order - End-to-end architecture - Compact 0.8B model - Apache-2.0 license For benchmark results, technical details, and training methodology, please visit the original repository. 2 --- # Credits ## Original Model All credit for the original model, datasets, training methodology, evaluation, benchmarks, and research belongs entirely to: **ATH-MaaS** Original Repository: https://huggingface.co/ATH-MaaS/OvisOCR2 Technical Report: https://arxiv.org/abs/2607.13639 --- ## oQe6 MLX Quantized Release This repository provides an Apple Silicon optimized **oQe6 MLX quantized** version of the original model. No additional fine-tuning has been performed. --- # Acknowledgements - ATH-MaaS - Qwen Team - Apple MLX - Hugging Face - Transformers - vLLM - SGLang - MLX-VLM - oMLX - OptiQ Quantization --- # Citation If you use this model in research, please cite the original OvisOCR2 paper: ```bibtex @misc{lu2026ovisocr2, title = {OvisOCR2 Technical Report}, author = {Shiyin Lu and Yinglun Li and Yu Xia and Yuhui Chen and An-Yang Ji and Jun-Peng Jiang and Qing-Guo Chen and Jianshan Zhao and En Lin and Haijun Li and Cheng Qin and Zhao Xu and Weihua Luo}, year = {2026}, eprint = {2607.13639}, archivePrefix= {arXiv}, primaryClass = {cs.CV} } ``` --- # License This release inherits the **Apache-2.0** license from the original model. Please refer to the original repository for complete licensing information. --- # Disclaimer This repository contains an optimized **oQe6 MLX quantized conversion** intended for efficient local inference on Apple Silicon. All original model architecture, datasets, training methodology, benchmarks, evaluations, and research remain entirely the work of the original ATH-MaaS team. ````3