Xiaomi-OCR-0 Technical Report

Xin Chen; Anan Du; Feng Feng; Pei Fu; Jian Luan; Longwei Xu; Shaojie Zhang; Hang Li

cs.CVSep 28, 2026

Xiaomi-OCR-0 Technical Report

Authors: Xin Chen, Anan Du, Feng Feng, Pei Fu, Jian Luan, Longwei Xu, Shaojie Zhang, Hang Li, +2 more

Organizations: SeerRay Team

Abstract

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.

Figures & tables

Figure 2 : Automated annotation pipeline. Heterogeneous experts annotate document regions, triplet consensus scores them, and render-based verification refines them when agreement is insufficient. Refined candidates return to the expert pool for re-consensus.

Figure 3 : Overview of the progressive training recipe.

Method	Size	Overall $\uparrow$	Text Edit $\downarrow$	Formula CDM $\uparrow$	Table TEDS $\uparrow$	Table TEDS-S $\uparrow$	Reading Order Edit $\downarrow$
General VLMs
Ovis2.6-30B-A3B [ 42 ]	30B	93.70	0.035	95.17	89.44	92.40	0.135
Gemini 3 Pro [ 25 ]	-	92.91	0.064	95.99	89.15	92.96	0.165
Gemini 3 Flash [ 18 ]	-	92.62	0.066	95.16	89.29	93.51	0.172
Qwen3-VL-235B [ 2 ]	235B	89.78	0.063	92.55	83.07	86.75	0.166
GPT-5.2 [ 52 ]	-	86.59	0.114	88.21	82.95	87.93	0.193

Table 1 : Comparison on OmniDocBench v1.6 [ 65 ] .

Method	Size	Overall $\uparrow$	Scanning $\uparrow$	Warping $\uparrow$	Screen-Photo $\uparrow$	Illumination $\uparrow$	Skew $\uparrow$
General VLMs
Kimi-K2.6 [ 49 ]	1T	89.76	90.08	89.62	89.58	89.91	89.61
Gemini 3 Pro [ 25 ]	-	89.24	89.47	88.90	88.86	89.53	89.45
Kimi K2.5 [ 33 ]	1T	89.09	89.67	88.86	88.39	89.66	88.86
Doubao-Seed-2.1-Pro [ 5 ]	-	89.02	88.85	89.36	88.99	89.13	88.79
Qwen3-VL-235B [ 2 ]	235B	88.90	89.43	89.99	89.27	89.27	86.56

Table 2 : Comparison on Real5-OmniDocBench [ 82 ] .

Method	Size	Overall $\uparrow$	Text Edit $\downarrow$	Formula CDM $\uparrow$	Table TEDS $\uparrow$	Table TEDS-S $\uparrow$	Reading Order Edit $\downarrow$
Specialized OCR Models
Logics-Parsing-v2 [ 10 ]	4B	77.10	0.4029	91.40	80.19	87.16	0.2355
HunyuanOCR-1.5 [ 36 ]	1B	77.62	0.1979	85.12	67.54	70.67	0.2750
dots.ocr [ 38 ]	3B	81.84	0.1483	85.00	75.32	80.20	0.2200
PaddleOCR-VL-1.5 [ 14 ]	0.9B	84.64	0.1461	86.72	81.80	86.52	0.2138
GLM-OCR [ 20 ]	0.9B	85.08	0.1514	89.09	81.31	85.90	0.2228

Table 3 : Comparison on Wild-OmniDocBench [ 34 ] .

Table 6

Model	Size	DocVQA $\uparrow$	InfoVQA $\uparrow$	ChartQA $\uparrow$	OCRBench $\uparrow$	TextVQA $\uparrow$	Overall $\uparrow$
General VLMs
GPT-5.2 [ 52 ]	-	91.7	84.0	57.0	80.7	72.8	77.2
GLM-4.5V [ 62 ]	106B-A12B	94.5	84.1	86.6	87.2	72.0	84.9
Gemini 3 Pro [ 25 ]	-	–	–	57.2	94.0	–	–
Qwen3.5-0.8B [ 55 ]	0.8B	88.5	60.3	69.5	77.9	68.3	72.9
Qwen3.5-2B [ 55 ]	2B	92.4	72.4	77.0	85.9	76.9	80.9

Table 6 : Comparison on document-oriented visual question answering benchmarks. Overall is the arithmetic mean of DocVQA, InfoVQA, ChartQA, OCRBench, and TextVQA on a 0–100 scale; it is reported only when all five scores are available. “–” denotes unavailable results.

Model Type	Model	Model Size	Average
Open-source General VLMs	Gemma 4 31B it [ 61 ]	31B	0.35
	InternVL3.5-8B [ 67 ]	8B	0.39
	MiniCPM-V 4.5 [ 77 ]	8B	0.40
	GLM-4.5V [ 62 ]	106B-A12B	0.43
	Ovis2.6-30B-A3B [ 42 ]	30B-A3B	0.51
	InternVL3.5-A28B [ 67 ]	241B-A28B	0.56

Table 7 : Comparison on mature-script (Clerical, Regular, Running, and Cursive) subset of Chronicles-OCR [ 35 ] .

Table 8 : Task-level key information extraction accuracy (%) on our in-house benchmark.

Training data	Overall $\uparrow$	Text Edit $\downarrow$	Formula CDM $\uparrow$	Table TEDS $\uparrow$	Table TEDS-S $\uparrow$	Reading Order Edit $\downarrow$
Early stage: 9% of all parsing data
Region-level parsing	93.594	0.045	96.555	88.727	92.237	0.144
+ OCR-VQA	93.264	0.047	96.368	88.123	91.538	0.145
Middle stage: 29% of all parsing data
Without page-level data	94.730	0.045	96.811	91.878	94.574	0.143
+ Page-level parsing	94.941	0.043	97.044	92.079	94.819	0.142

Table 10 : Understanding-to-parsing transfer at three stages of parsing training on OmniDocBench v1.6. Each block shows the successive addition of parsing and OCR-VQA supervision.

Checkpoint	Overall $\uparrow$	Text Edit $\downarrow$	Formula CDM $\uparrow$	Table TEDS $\uparrow$	Table TEDS-S $\uparrow$	Reading Order Edit $\downarrow$
baseline	72.21	0.186	70.728	64.491	70.010	0.179
Q-Mask ablation
CPT without Q-Mask	96.0241	0.0378	97.5249	94.3273	96.8657	0.1396
+ Q-Mask	96.3198	0.0372	98.0180	94.6613	96.9934	0.1395
Mix-RL ablation
CPT starting checkpoint	96.4739	0.0332	98.3444	94.3973	96.6835	0.1221

Table 11 : Component ablations on OmniDocBench v1.6, with the 0.8B baseline for reference. The Q-Mask and Mix-RL comparisons are evaluated at different training stages.

(a) Document parsing: OmniDocBench v1.6
Model	Overall $\uparrow$	Text Edit $\downarrow$	Formula CDM $\uparrow$	Table TEDS $\uparrow$	Table TEDS-S $\uparrow$	Reading Order Edit $\downarrow$
4B parsing teacher	96.9745	0.0308	98.4102	95.5932	97.5050	0.1228
0.8B CPT	96.4739	0.0332	98.3444	94.3973	96.6835	0.1221
0.8B MOPD	96.7417	0.0314	98.4677	94.8975	97.1579	0.1223
0.8B Mix-RL	96.8277	0.0313	98.5044	95.1087	97.1929	0.1221

Table 12 : Task-specific 4B teachers and 0.8B students on parsing and OCR-centric understanding. The Mix-RL rows report the same model results as the Ours rows in Tables 1 and 6 , respectively. VQA Overall is the arithmetic mean of all five benchmarks on a 0–100 scale. The best value in each column is in bold.

Figure 4 : Training curves for Mix-RL and MOPD on document parsing and OCR-centric understanding. Both panels report Overall scores; the mean over OCRBench, DocVQA, TextVQA, InfoVQA, and ChartQA measures understanding. Horizontal position indicates cumulative training compute on a shared relative scale.

Explore similar work

OvisOCR2 Technical Report
Jul 15, 2026Shiyin Lu, Yinglun Li, Yu Xia +10Document Parsing Technical Report
ATH-MaaS, Alibaba Group
PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training
Jun 2, 2026Zelun Zhang, Hongen Liu, Suyin Liang +12Document Post-Training
PaddlePaddle Team, Baidu Inc.
MonkeyOCRv2: A Visual-Text Foundation Model for Document AI
Jul 13, 2026Yuliang Liu, Zhang Li, Ziyang Zhang +11Lingdt-Vl-Ocr Visual Document Retrieval
1Huazhong University of Science and Technology · 2Kingsoft Office

Jul 15, 2026, cs.CVJ/K move · Enter open · S save

OvisOCR2 Technical Report

Shiyin Lu, Yinglun Li, Yu Xia, Yuhui Chen+9

ATH-MaaS, Alibaba Group

Document Parsing Technical Report Report

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

OvisOCR2 Technical Report

PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training

MonkeyOCRv2: A Visual-Text Foundation Model for Document AI