Document parsing is a fine-grained task where image resolution significantly impacts performance. While advanced research leveraging vision-language models benefits from high-resolution input to boost model performance, this often leads to a quadratic increase in the number of vision tokens and significantly raises computational costs. We attribute this inefficiency to substantial visual regions redundancy in document images, like background. To tackle this, we propose PaddleOCR-VL, a novel coarse-to-fine architecture that focuses on semantically relevant regions while suppressing redundant ones, thereby improving both efficiency and performance. Specifically, we introduce a lightweight Valid Region Focus Module (VRFM) which leverages localization and contextual relationship prediction capabilities to identify valid vision tokens. Subsequently, we design and train a compact yet powerful 0.9B vision-language model (PaddleOCR-VL-0.9B) to perform detailed recognition, guided by VRFM outputs to avoid direct processing of the entire large image. Extensive experiments demonstrate that PaddleOCR-VL achieves state-of-the-art performance in both page-level parsing and element-level recognition. It significantly outperforms existing solutions, exhibits strong competitiveness against top-tier VLMs, and delivers fast inference while utilizing substantially fewer vision tokens and parameters, highlighting the effectiveness of targeted coarse-to-fine parsing for accurate and efficient document understanding. The source code and models are publicly available at https://github.com/PaddlePaddle/PaddleOCR.
Figures & tables
Figure 1: PaddleOCR-VL achieves state-of-the-art performance with the fewest vision tokens and parameters on OmniDocBench v1.5.
Figure 2: Architectural comparison of End-to-end VLM and our method. Among various types of document images, the valid area accounts for less than 50% on average. Therefore, our proposed PaddleOCR-VL can achieve improvements in visual information processing by discarding redundant regions.
Figure 3: The overview of our proposed PaddleOCR-VL, which consists of two components. The first component is VRFM, which accurately extracts valid regions containing layout elements from unstructured document images and predicts the reading order. Subsequently, each valid region is fed into our designed PaddleOCR-VL-0.9B for precise recognition of individual elements. The final results are then reconstructed into a structured document according to the reading order.
Figure 4: Architecture of the Valid Region Focus Module (VRFM).
Stages
Stage 1
Stage 2
Training Samples
29M
2.7M
Max Resolution
1280 × 28 × 28
2048 × 28 × 28
Sequence length
16384
16384
Trainable components
All
All
Batch sizes
128
128
Data Augmentation
Yes
Yes
Table 1: Training settings of PaddleOCR-VL-0.9B.
Model Type
Methods
Parameters
Vision Tokens
Overall ↑
Text Edit ↓
Formula CDM ↑
Table TEDS ↑
Table TEDS-S ↑
Reading Order Edit ↓
Pipeline Tools
Marker-1.8.2 [ 44 ]
-
-
71.30
0.206
76.66
57.88
71.17
0.250
Mineru2-pipeline [ 40 ]
-
-
75.51
0.209
76.55
70.90
79.11
0.225
PP-StructureV3 [ 9 ]
-
-
86.73
0.073
85.79
81.68
89.48
0.073
General VLMs
GPT-4o [ 1 ]
-
-
75.02
0.217
79.70
67.07
76.09
0.148
InternVL3-76B [ 63 ]
76B
7277
80.33
0.131
83.42
70.64
77.74
0.113
InternVL3.5-241B [ 53 ]
241B
7277
82.67
0.142
87.23
75.00
81.28
0.125
Table 2: Comprehensive evaluation of document parsing on OmniDocBench v1.5. Results are reported by OmniDocBench [ 41 ] unless Ours. In PaddleOCR-VL, S, M, and L refer to the same PaddleOCR-VL weights using different vsion tokens.
Methods
Overall TEDS ↑
Structural TEDS ↑
Overall Edit Dist ↓
MinerU2-VLM [ 40 ]
0.9002
0.9369
0.0734
dots.ocr [ 46 ]
0.8194
0.8442
0.1508
MinerU2.5 [ 39 ]
0.9005
0.9539
0.0693
PaddleOCR-VL-L
0.9046
0.9420
0.0681
Table 3: Comparison of OmniDocBench-Table-block Performance
Methods
Edit Distance ↓
PPT
Academic Literature
Book
Colorful Textbook
Exam Paper
Magazine
Newspaper
Note
Research Report
Qwen2.5-VL-72B [ 3 ]
0.054
0.023
0.061
0.084
0.195
0.032
0.056
0.118
0.040
MonkeyOCR-pro-3B [ 23 ]
0.058
0.021
0.064
0.096
0.116
0.023
0.058
0.124
0.052
MinerU2.5 [ 39 ]
0.195
0.089
0.111
0.234
0.194
0.147
0.056
0.142
0.094
Dolphin [ 13 ]
0.237
0.095
0.135
0.347
0.248
0.233
0.121
0.309
0.213
PaddleOCR-VL-L
0.049
0.021
0.047
0.082
0.115
0.020
0.035
0.077
0.031
Table 4: Overall Comparison of OmniDocBench-OCR-block Performance.
Methods
Overall CDM ↑
EN CDM ↑
ZH CDM ↑
dots.ocr [ 46 ]
0.4641
0.4868
0.4414
MinerU2-VLM [ 40 ]
0.8286
0.9616
0.6956
MonkeyOCR-pro-1.2B [ 23 ]
0.8531
0.9642
0.7419
MonkeyOCR-3B [ 23 ]
0.8621
0.9718
0.7524
Qwen2.5-VL-72B [ 3 ]
0.8747
0.9574
0.7920
MinerU2.5 [ 39 ]
0.9187
0.9751
0.8623
Table 5: Comparison of OmniDocBench v1.5 Formula-block Performance. Due to dots.ocr [ 46 ] easily recognizing cropped formulas as images, the score is relatively low.
Methods
RMS-F1 ↑
Overall
EN
ZH
TinyChart [ 59 ]
0.2159
0.4726
0.0876
GOT [ 54 ]
0.3160
0.1100
0.4190
OneChart [ 7 ]
0.3716
0.1384
0.4882
Qwen2.5-VL-72B [ 3 ]
0.7300
0.6972
0.7464
PP-StructureV3 [ 9 ]
0.8060
0.7963
0.8109
Table 6: Comparison of In-house-Chart Performance
Methods
Total Time (s) ↓
Pages/s ↑
Tokens/s ↑
Avg. VRAM Usage (GB) ↓
MonkeyOCR-pro-1.2B [ 23 ]
1456.4
0.6730
1120.3
75.5
dots.ocr [ 46 ]
2784.6
0.3522
532.9
78.5
MinerU2.5 [ 39 ]
927.3
1.0574
1647.9
41.9
PaddleOCR-VL-L
605.2
1.6192
2470.7
42.1
Table 7: End-to-End Inference Performance Comparison. All models were tested using the vLLM 0.10.2 backend.
Figure S1 : The construction method and characteristics of the text training data for PaddleOCR-VL-0.9B.
Figure S2 : The construction method and characteristics of the table training data for PaddleOCR-VL-0.9B.
Figure S3 : The construction method and characteristics of the formula training data for PaddleOCR-VL-0.9B.
Figure S4 : The construction method and characteristics of the chart training data for PaddleOCR-VL-0.9B.
Language Category
Specific Languages
Chinese
Chinese
English
English
Korean
Korean
Japanese
Japanese
Thai
Thai
Greek
Greek
Table S1 : Supported Languages
Method
Visual Tokens
Acc.(%)
Qwen3-VL-8B [ 2 ]
5818
89.60
DocKylin [ 58 ] + Qwen3-VL-8B [ 2 ]
4316
87.75
PP-DocLayout [ 47 ] + Qwen3-VL-8B
3876
89.30
PP-DocLayout + PaddleOCR-VL-0.9B
2347
89.96
VRFM + Qwen3-VL-8B
4262
90.37
VRFM + PaddleOCR-VL-0.9B
2561
92.62
Table S2: Comparison of VRFM with representative baselines in terms of token efficiency and accuracy.
We introduce PaddleOCR-VL-1.6, an upgraded compact document parsing model built upon PaddleOCR-VL-1.5. Although PaddleOCR-VL-1.5 establishes a strong 0.9B baseline, its remaining errors concentrate in under-optimized regions where model behavior is unstable, data coverage is sparse, or supervision is unreliable. Rather than expanding the training corpus indiscriminately, PaddleOCR-VL-1.6 introduces a region-aware data optimization framework that identifies weak regions from the previous model, applies targeted enhancement to these regions, and improves the reliability of supervision signals. It further adopts a progressive post-training recipe based on curated data selection and reinforcement learning, pushing model performance to a higher level through staged optimization. PaddleOCR-VL-1.6 achieves a new state-of-the-art score of 96.33% on OmniDocBench v1.6, demonstrates strong competitiveness against top-tier VLMs, and provides a practical post-training recipe for the PaddleOCR-VL series.
Document parsing, as a fundamental yet crucial vision task, is being revolutionized by vision-language models (VLMs). However, the autoregressive (AR) decoding inherent to VLMs creates a significant bottleneck, severely limiting parsing speed. In this paper, we propose Parallel-Token Prediction (PTP), a plugable, model-agnostic and simple-yet-effective method that enables VLMs to generate multiple future tokens in parallel with improved sample efficiency. Specifically, we insert some learnable tokens into the input sequence and design corresponding training objectives to equip the model with parallel decoding capabilities for document parsing. Furthermore, to support effective training, we develop a comprehensive data generation pipeline that efficiently produces large-scale, high-quality document parsing training data for VLMs. Extensive experiments on OmniDocBench and olmOCR-bench demonstrate that our method not only significantly improves decoding speed (1.6x-2.2x) but also reduces model hallucinations and exhibits strong generalization abilities.
Lei Li, Ze Zhao, Meng Li +6
Platform and Content Group, Tencent · Renmin University of China
Document parsing is a fundamental task in multimodal understanding, supporting a wide range of downstream applications such as information extraction and intelligent document analysis. Benefiting from strong semantic modeling and robust generalization, VLM-based end-to-end approaches have emerged as the mainstream paradigm in recent years. However, these models often suffer from substantial inference latency, as they must autoregressively generate long, full-page sequences when processing long-form documents. While recent hybrid methods mitigate this issue via region-level parallel decoding with VLMs, independent region decoding loses full-page context and might weaken global coherence. To address this issue, we propose Hierarchical Speculative Decoding (HSD), a two-stage local-to-global framework for document parsing. HSD first employs a lightweight pipeline drafter to predict region partitions and generate coarse drafts for each region. The first stage verifies the generated region-level drafts in parallel for efficiency, while the second stage further performs page-level verification on these refined outputs to preserve full-page coherence. Experimental results show that HSD achieves a near-lossless 2.7x speedup with HunyuanOCR on OmniDocBench v1.5 and up to 7.04x speedup on long-document parsing tasks, demonstrating the effectiveness of the proposed method. The code is available at https://github.com/whlscut/HSD.
Wenhui Liao, Hongliang Li, Pengyu Xie +15
South China University of Technology, Guangzhou, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China · Nanjing University, Nanjing, China +1