OvisOCR2 F32 GGUF
> OvisOCR2, developed by the ATH-MaaS team, is a compact 0.8B vision-language model for end-to-end page-level document parsing, built by post-training Qwen3.5-0.8B to directly convert document images into structured...
Fuente del modelo
- Tipo
- Modelo
- Tamaño del archivo
- 111 MB–2,81 GB
- Modelo base
- ATH-MaaS/OvisOCR2
- Versión
- vRevision da14f83
- Actualizado
- 30 oct 2026
- Descargas
- 738
- Fuente
- Hugging Face
Descripción de la fuente
OvisOCR2-F32-GGUF
OvisOCR2, developed by the ATH-MaaS team, is a compact 0.8B vision-language model for end-to-end page-level document parsing, built by post-training Qwen3.5-0.8B to directly convert document images into structured Markdown while preserving the natural reading order. It accurately extracts plain text, complex tables (as HTML), mathematical formulas (as LaTeX), and visual regions, making it suitable for OCR, document understanding, and digitization workflows. Despite its small size, OvisOCR2 achieves state-of-the-art performance with an overall score of 96.58 on OmniDocBench v1.6, becoming the first end-to-end model to surpass traditional pipeline-based approaches on the benchmark, while also leading PureDocBench with an Avg3 score of 75.06. Trained using a combination of real-world and synthetic data with a multi-stage recipe incorporating Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), and On-Policy Distillation (OPD), OvisOCR2 delivers high-quality document parsing with a lightweight deployment footprint and native support for efficient inference through vLLM.
Model Files
| File Name | Quant Type | File Size | File Link |
|---|---|---|---|
| OvisOCR2.BF16.gguf | BF16 | 1.52 GB | Download |
| OvisOCR2.F16.gguf | F16 | 1.52 GB | Download |
| OvisOCR2.F32.gguf | F32 | 3.02 GB | Download |
| OvisOCR2.Q3_K_L.gguf | Q3_K_L | 491 MB | Download |
| OvisOCR2.Q3_K_M.gguf | Q3_K_M | 466 MB | Download |
| OvisOCR2.Q3_K_S.gguf | Q3_K_S | 435 MB | Download |
| OvisOCR2.Q4_K_M.gguf | Q4_K_M | 529 MB | Download |
| OvisOCR2.Q4_K_S.gguf | Q4_K_S | 505 MB | Download |
| OvisOCR2.Q5_K_M.gguf | Q5_K_M | 578 MB | Download |
| OvisOCR2.Q5_K_S.gguf | Q5_K_S | 564 MB | Download |
| OvisOCR2.Q8_0.gguf | Q8_0 | 812 MB | Download |
| OvisOCR2.mmproj-bf16.gguf | mmproj-bf16 | 207 MB | Download |
| OvisOCR2.mmproj-f16.gguf | mmproj-f16 | 207 MB | Download |
| OvisOCR2.mmproj-q8_0.gguf | mmproj-q8_0 | 116 MB | Download |
llama.cpp
LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp
Fuentes
1 fuente
Fuentes
1 fuenteVerificado 30 oct
Download files
Repository files
File inventory only; completeness is unverified.
Artefactos del modelo
14 artefactos
Artefactos del modelo
14 artefactosExtractos de fuentes
3 extractos
Extractos de fuentes
3 extractos--- tags: - text-generation-inference - ocr - document-parsing - multimodal - markdown - tables - formulas - vllm - qwen3_5 license: apache-2.0 language: - en base_model: - ATH-MaaS/OvisOCR2 pipeline_tag: image-text-to-text library_name: transformers --- # **OvisOCR2-F32-GGUF** > **OvisOCR2**, developed by the **[ATH-MaaS](https://huggingface.co/ATH-MaaS)** team, is a compact **0.8B** vision-language model for **end-to-end page-level document parsing**, built by post-training **Qwen3.5-0.8B** to directly convert document images into structured **Markdown** while preserving the natural reading order. It accurately extracts plain text, complex tables (as HTML), mathematical formulas (as LaTeX), and visual regions, making it suitable for OCR, document understanding, and digitization workflows. Despite its small size, OvisOCR2 achieves **state-of-the-art performance** with an overall score of **96.58** on **OmniDocBench v1.6**, becoming the first end-to-end model to surpass traditional pipeline-based approaches on the benchmark, while also leading **PureDocBench** with an **Avg3 score of 75.06**. Trained using a combination of real-world and synthetic data with a multi-stage recipe incorporating **Supervised Fine-Tuning (SFT)**, **Reinforcement Learning (RL)**, and **On-Policy Distillation (OPD)**, OvisOCR2 delivers high-quality document parsing with a lightweight deployment footprint and native support for efficient inference through **vLLM**. ## Model Files File Name | Quant Type | File Size | File Link | |-----------|------------|-----------|-----------| | OvisOCR2.BF16.gguf | BF16 | 1.52 GB | [Download](https://huggingface.co/prithivMLmods/OvisOCR2-F32-GGUF/blob/main/OvisOCR2.BF16.gguf) | | OvisOCR2.F16.gguf | F16 | 1.52 GB | [Download](https://huggingface.co/prithivMLmods/OvisOCR2-F32-GGUF/blob/main/OvisOCR2.F16.gguf) | | OvisOCR2.F32.gguf | F32 | 3.02 GB | [Download](https://huggingface.co/prithivMLmods/OvisOCR2-F32-GGUF/blob/main/OvisOCR2.F32.gguf) | | OvisOCR2.Q3_K_L.gguf | Q3_K_L | 491 MB | [Download](https://huggingface.co/prithivMLmods/OvisOCR2-F32-GGUF/blob/main/OvisOCR2.Q3_K_L.gguf) | | OvisOCR2.Q3_K_M.gguf | Q3_K_M | 466 MB | [Download](https://huggingface.co/prithivMLmods/OvisOCR2-F32-GGUF/blob/main/OvisOCR2.Q3_K_M.gguf) | | OvisOCR2.Q3_K_S.gguf | Q3_K_S | 435 MB | [Download](https://huggingface.co/prithivMLmods/OvisOCR2-F32-GGUF/bl...
Source context: 738 downloads · 4 likes · Pipeline image-text-to-text · Library transformers · Repo prithivMLmods/OvisOCR2-F32-GGUF
Source context: 3689 downloads · 3 likes · Pipeline image-text-to-text · Library transformers · Repo prithivMLmods/OvisOCR2-F32-GGUF
