We introduce OvisOCR3, a 0.8B model for a broad range of OCR-related tasks. OvisOCR3 unifies document parsing, general OCR, key information extraction (KIE), document figure captioning, and cross-page table reconstruction in a single model. We train the model on real and synthetic data through continued pretraining and use domain-balanced mid-training to balance supervision across tasks with different data volumes and response lengths. Multi-teacher on-policy distillation transfers expertise from five domain-specific 4B teachers into the shared model, followed by mixed-task reinforcement learning with domain-specific rewards. We compare OvisOCR3 with general-purpose vision-language models and specialized OCR models on public and in-house benchmarks covering all five task families. In these comparisons, it achieves an overall score of 96.74 on OmniDocBench v1.6 and the highest overall scores on PureDocBench, Real5-OmniDocBench, and MDPBench. On OCRBench v2, it ranks first, scoring 65.9 and 61.3 on the English and Chinese splits, respectively. It also outperforms the baselines on OCRBench and all four KIE benchmarks. On DocFigCapBench, it achieves the highest localization F1 and a caption score close to that of Qwen3.8-Max. For cross-page tables, it leads in Relation F1 on CrossTabBench and truncated-table TEDS on MPDocBench-Parse. These results demonstrate that a compact 0.8B model can achieve strong performance across diverse OCR-related tasks. OvisOCR3 is available at https://huggingface.co/ATH-MaaS/OvisOCR3.
Figures & tables
Figure 1: Performance of OvisOCR3 across benchmarks for OCR-related tasks.
Task family
Task scope
Output representation
Document parsing
Recovery of page content, layout, and reading order
Region-level JSON or Markdown with LaTeX and HTML
General OCR
Recognition, localization, translation, and visual reasoning
Transcriptions, bounding boxes, LaTeX, chart data, labels, or answers
Key information extraction
Extraction and organization of fields from images
Key-value records or template-guided JSON
Document figure captioning
Localization and description of document figures
Localized captions and Mermaid representations
Cross-page table reconstruction
Continuation detection and table reconstruction across pages
HTML tables with source-page associations, or a no-continuation prediction
Table 1: Task coverage and output representations of OvisOCR3.
Task
Additional input
Output
Plain OCR
None
Plain-text transcription
Word-level text spotting
None
JSON records with word boxes and text
Reference grounding
Ordered reference list
JSON records with reference boxes
Region OCR
Target bounding box
Text within the region
Formula recognition
None
LaTeX expressions
Chart parsing
None
Chart metadata and nested data values
Table 2: The ten General OCR interfaces. Every task receives an image; the second column lists additional task inputs. Bounding boxes use the order [xmin,ymin,xmax,ymax] , with integer coordinates normalized independently by image width and height to [0,1000) .
Figure 2: Two-stage cross-page table reconstruction with a shared model. Page-level parsing produces Markdown with HTML tables; reconstruction uses these representations together with adjacent page images to identify continuations and recover complete tables.
Figure 3: Recursive reconstruction of a table spanning three pages. After merging the fragments on pages 2 and 3, the updated Markdown for page 2 is re-rendered and paired with page 1. Colors denote the source pages of the table fragments. Each step uses two page images and their current parsed representations.
Model Type
Model
Parameters
Overall ↑
Text Edit ↓
Formula CDM ↑
Table TEDS ↑
Table TEDS-S ↑
RO Edit ↓
General VLMs
InternVL3.5-241B [ 19 ]
241B/28B
83.76
0.1300
89.95
74.35
79.78
0.2150
Kimi K2.5 [ 20 ]
1T/32B
84.53
0.1070
83.50
80.76
84.00
0.2110
GPT-5.2 [ 21 ]
–
86.59
0.1140
88.21
82.95
87.93
0.1930
Qwen3-VL-235B [ 22 ]
235B/22B
89.78
0.0630
92.55
83.07
86.75
0.1660
Gemini 3 Flash [ 23 ]
–
92.62
0.0660
95.16
89.29
93.51
0.1720
Gemini 3 Pro [ 24 ]
–
92.91
0.0640
95.99
89.15
92.96
0.1650
Table 3: Document parsing results on OmniDocBench v1.6. Baseline scores without † are taken from the official leaderboard [ 15 ] and model reports [ 4 , 16 , 17 , 18 ] . RO denotes reading order.
Model Type
Model
Parameters
Overall ↑
Clean ↑
Digital ↑
Real ↑
General VLMs
GPT-5.6 Sol [ 42 ]
–
69.16
74.15
71.73
61.59
Kimi K2.6 [ 43 ]
1T/32B
70.10
72.32
69.95
68.02
Qwen3.5-122B-A10B [ 44 ]
122B/10B
74.11
76.14
76.34
69.85
Claude Opus 5 [ 45 ]
–
77.07
82.99
72.94
75.27
Qwen3.8-27B [ 46 ]
27B
77.63
80.45
78.86
73.59
Gemini 3.6 Flash [ 47 ]
–
78.27
79.97
79.11
75.73
Table 4: Document parsing results on PureDocBench. Overall averages the Clean, Digital, and Real track scores. Baseline scores without † follow the benchmark leaderboard [ 41 ] .
Model Type
Model
Parameters
Overall ↑
Scanning ↑
Warping ↑
Screen Photo. ↑
Illum. ↑
Skew ↑
General VLMs
GPT-5.2 [ 21 ]
–
78.66
84.43
76.26
76.75
80.88
75.00
Qwen3.8-Max † [ 25 ]
–
86.86
91.80
90.58
80.59
86.18
85.17
Qwen2.5-VL-72B [ 54 ]
72B
86.92
86.19
87.77
86.48
87.25
86.90
Qwen3-VL-235B [ 22 ]
235B/22B
88.90
89.43
89.99
89.27
89.27
86.56
Doubao-Seed-2.1-Pro [ 55 ]
–
89.02
88.85
89.36
88.99
89.13
88.79
Gemini 3 Pro [ 24 ]
–
89.24
89.47
88.90
88.86
89.53
89.45
Table 5: Document parsing results on Real5-OmniDocBench. Overall averages the scores across the five acquisition conditions; Illum. denotes illumination. Baseline scores without † follow the official leaderboard [ 53 ] .
Model Type
Model
Parameters
Overall ↑
Digital ↑
Photo ↑
Latin ↑
Non-Latin ↑
General VLMs
Qwen3-VL-8B [ 22 ]
8B
68.3
78.4
65.0
73.6
62.5
GPT-5.2 [ 21 ]
–
68.6
85.6
63.0
75.2
61.1
Claude Sonnet 4.6 [ 58 ]
–
73.1
85.0
69.3
79.2
66.2
Doubao-Seed-2.0-Pro [ 59 ]
–
74.2
78.9
72.8
75.7
72.5
Kimi K3 [ 60 ]
2.8T/104B
83.6
90.8
81.2
86.2
80.7
Gemini 3 Pro [ 24 ]
–
86.4
90.4
85.1
88.4
84.1
Table 6: Document parsing results on the MDPBench public split, with breakdowns by acquisition type and script group. Baseline scores without † follow the official leaderboard [ 57 ] .
Model
Parameters
OCRBench
v2 English
v2 Chinese
Qwen3.5-0.8B [ 5 ]
0.8B
79.1
41.5
43.9
Qwen3.5-2B [ 66 ]
2B
85.4
48.1
55.4
Qwen3.5-4B [ 6 ]
4B
85.0
54.1
61.0
Qwen3.5-OCR [ 30 ]
–
76.8
44.5
43.8
PaddleOCR-VL-1.6 [ 39 ]
0.9B
55.9
23.8
30.4
HunyuanOCR-1.5 [ 16 ]
1B
86.1
50.7
57.7
Table 7: Performance comparison on OCRBench and OCRBench v2.
Task category
Samples
Accepted
Score (%)
Text Recognition
300
284
94.7
Scene Text-centric VQA
200
167
83.5
Doc-oriented VQA
200
180
90.0
Key Information Extraction
200
181
90.5
Handwritten Mathematical Expression Recognition
100
76
76.0
Total
1,000
888
88.8
Table 8: Task-level performance of OvisOCR3 on OCRBench.
Task category
Language
Samples
Score
Text recognition
EN
1,200
76.9
Text detection
EN
500
66.5
Text spotting
EN
200
69.9
Relationship extraction
EN
700
87.7
Element parsing
EN
1,600
59.2
Mathematical calculation
EN
500
34.6
Table 9: Task-level performance of OvisOCR3 on OCRBench v2.
Model
OCRBench KIE
OCRBench v2 KIE
CC-OCR KIE
In-house KIE Bench
Qwen3.5-OCR [ 30 ]
78.50
72.90
88.43
59.90
GLM-OCR [ 37 ]
79.00
74.22
68.36
75.25
HunyuanOCR-1.5 [ 16 ]
81.00
68.63
66.08
75.96
Xiaomi-OCR-0 [ 18 ]
76.50
74.66
79.53
75.68
OvisOCR3
90.50
83.01
88.69
77.73
Table 10: Performance comparison on KIE benchmarks.
Figure 4: An example document page from DocFigCapBench. (a) The full page, with two figure targets in red and other page annotations in gray. (b) Reference figure captions and a summary of the remaining annotations.
Figure 5: An example of a four-image figure group from DocFigCapBench. The red outline marks one evaluation target; the blue outlines mark its four annotated members.
Category
Model
Localization F1 ↑
Caption Score ↑
General VLMs
Qwen3.5-0.8B [ 5 ]
5.47
0.20
Qwen3.5-2B [ 66 ]
10.09
0.42
Qwen3.5-4B [ 6 ]
40.19
2.27
Qwen3.7-Plus [ 69 ]
71.14
5.69
Qwen3.8-Max [ 25 ]
86.67
7.30
Specialized OCR Models
HunyuanOCR-1.5 [ 16 ]
48.24
–
Table 11: Performance comparison on DocFigCapBench.
Figure 6: An annotated two-page example from CrossTabBench. Red boxes mark fragments of the same table, with enlarged views below the source pages. The ground-truth reconstruction on the right preserves merged cells and numerical values.
CrossTabBench
MPDocBench-Parse
Model
Relation F1 (P/R)
Reconstruction TEDS
Truncated Table TEDS
PaddleOCR-VL-1.6 [ 39 ]
42.61 (90.13/27.90)
24.24
83.39
Logics-Parsing-V3 [ 70 ]
44.55 (90.63/29.53)
26.10
88.05
MinerU2.5-Pro [ 38 ]
59.14 (93.81/43.18)
41.02
91.44
MinerU2.5-Pro + MinerU-PoPo [ 38 , 71 ]
54.76 (97.40/38.09)
36.42
91.87
Qwen3.8-Max [ 25 ]
94.47 (90.19/99.19)
91.71
89.57
Table 12: Performance comparison on cross-page table reconstruction benchmarks.
Appendix figures & tables30 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A.1 : Structured document parsing with predicted region boundaries, semantic categories, and rendered page content.
Figure A.3 : Markdown document parsing, with text in reading order and mathematical expressions rendered from LaTeX.
Figure A.4 : Markdown parsing of a photographed document with page deformation and uneven illumination.
Figure A.5 : Markdown parsing of a photographed document affected by low light and perspective distortion.
Figure A.6 : Markdown document parsing of a Japanese page.
Figure A.7 : Plain OCR on a comic page, transcribing speech-balloon text in reading order.
Figure A.8 : Word-level text spotting. Each of the five recognized words is paired with a predicted bounding box.
Figure A.9 : Reference grounding for the supplied word AUTOMOBILE , returning its bounding box.
Figure A.10 : Region OCR on a specified paragraph, excluding surrounding text from the response.
Figure A.11 : Formula recognition with LaTeX output and its rendered matrix expression.
Figure A.12 : Chart parsing with metadata and per-category values organized as JSON. Numerical values are serialized as strings.
Figure A.13 : Text translation from Chinese to English within the selected region.
Figure A.14 : Image classification of ASCII art. The response selects basketball from the supplied candidate labels.
Figure A.15 : Visual question answering from a book cover, with a brief explanation followed by the author’s name as the final answer.
Figure A.16 : Position-aware VQA on a yellow sign. The response pairs the answer text with its predicted bounding box.
Figure A.17 : Open extraction from a CC-OCR [ 2 ] receipt. The model extracts item amounts and payment totals using keys drawn from the receipt, without a supplied field list or JSON template. Blurring is present in the source image.
Figure A.18 : Specified-field extraction from an OCRBench v2 [ 1 ] registration form. The response contains exactly the five requested keys. Missing applicant telephone and fax values remain empty, rather than being filled with the organizer’s contact details.
Figure A.19 : Schema-based extraction from a CC-OCR [ 2 ] receipt. Each item’s name, quantity, and price are grouped in one record; the total amount, cash payment, and change are stored in a separate nested object. Blurring is present in the source image.
Figure A.20 : Comparison on a page with multiple visual regions.
Figure A.21 : Comparison on a photographed magazine page.
Figure A.22 : Comparison on a page with two charts.
Figure A.23 : Comparison on a newspaper page with individual images and an image group.
Figure A.24 : OvisOCR3 generates a caption and a Mermaid representation for an algorithm flowchart. The other 14 records in the full-page response are omitted from this display.
Figure A.25 : OvisOCR3 generates a caption and a Mermaid representation for a device block diagram. The other 18 records in the full-page response are omitted from this display.
Figure A.26 : OvisOCR3 generates a caption and a Mermaid representation for a system architecture diagram. The other 45 records in the full-page response are omitted from this display.
Figure A.27 : Cross-page correlation table from MPDocBench-Parse (document pages 4–5). The FDI label spans two rows separated by the page boundary. Full pages and enlarged fragments show the source structure; Figure A.28 compares the reconstructions.
Figure A.28 : Reconstructions of the correlation table in Figure A.27 . MinerU2.5-Pro splits the shared FDI label cell, while Qwen3.8-Max misplaces and duplicates numerical values. Logics-Parsing-V3 omits the second-page fragment. OvisOCR3 preserves both the shared FDI cell and the placement of numerical values. Shading marks the rows spanning the page boundary.
Figure A.29 : Cross-page financial table from MPDocBench-Parse (document pages 6–7). The currency unit at the start of the second page completes the cash-flow label on the first page. Full pages and enlarged fragments show the source structure; Figure A.30 compares the reconstructions.
Figure A.30 : Reconstructions of the financial table in Figure A.29 . MinerU2.5-Pro , Qwen3.8-Max , and Logics-Parsing-V3 each place the currency unit in a separate row, leaving four numerical cells empty. OvisOCR3 attaches the unit to the preceding cash-flow label, avoiding the extra row. Shading marks the content split by the page boundary.
We introduce OvisOCR2, a 0.8B document parsing model. OvisOCR2 is designed as an end-to-end parser: given a document page image, it generates a Markdown representation in natural reading order, covering text, formulas, tables, and visual regions. We build a data engine that combines filtered real-document annotations with synthetic pages whose rendered images and Markdown targets are derived from the same HTML source. The training recipe includes supervised fine-tuning, reinforcement learning on a 4B branch with a multi-component reward design, on-policy distillation into the 0.8B model, and model fusion. On OmniDocBench v1.6, OvisOCR2 achieves a state-of-the-art overall score of 96.58, placing an end-to-end model at the top of this leaderboard previously dominated by pipeline methods and highlighting the potential of end-to-end document parsing. On PureDocBench, OvisOCR2 also achieves the highest Avg3 score of 75.06. Beyond these two public benchmarks, we evaluate OvisOCR2 on an in-house benchmark designed to cover a broader set of long-tail and challenging scenarios. OvisOCR2 obtains the best overall performance among the compared methods, providing further evidence of its generalization and robustness. OvisOCR2 is available at https://huggingface.co/ATH-MaaS/OvisOCR2.
We present Qianfan-OCR, a 4B-parameter end-to-end vision-language model that unifies document parsing, layout analysis, and document understanding within a single architecture. It performs direct image-to-Markdown conversion and supports diverse prompt-driven tasks including table extraction, chart understanding, document QA, and key information extraction. To address the loss of explicit layout analysis in end-to-end OCR, we propose Layout-as-Thought, an optional thinking phase triggered by special think tokens that generates structured layout representations -- bounding boxes, element types, and reading order -- before producing final outputs, recovering layout grounding capabilities while improving accuracy on complex layouts. Qianfan-OCR ranks first among end-to-end models on OmniDocBench v1.5 (93.12) and OlmOCR Bench (79.8), achieves competitive results on OCRBench, CCOCR, DocVQA, and ChartQA against general VLMs of comparable scale, and attains the highest average score on public key information extraction benchmarks, surpassing Gemini-3.1-Pro, Seed-2.0, and Qwen3-VL-235B. The model is publicly accessible via the Baidu AI Cloud Qianfan platform.
Compact OCR-specific vision-language models achieve strong document parsing performance, but often rely on costly supervision and focus primarily on visual-text reconstruction. We introduce Xiaomi-OCR-0, a unified 0.8B model for document parsing and OCR-centric understanding. We build an approximately 170M-sample OCR-centric corpus using an automated data engine that combines expert consensus, render-based verification, and targeted synthesis. Starting from Qwen3.5-0.8B, our progressive training recipe combines Q-Mask-based text anchoring, continued pretraining, and mixed-task reinforcement learning (Mix-RL). Xiaomi-OCR-0 achieves 95.24 on Real5-OmniDocBench, 96.83 on OmniDocBench v1.6, and 87.94 on Wild-OmniDocBench, while reaching an average score of 83.2 across five OCR-oriented VQA benchmarks. Ablations further show that, with sufficient parsing training, OCR-centric understanding supervision provides additional gains for document parsing. Homepage: https://huggingface.co/spaces/SeerRay-Lab/Xiaomi-OCR-0.