DeepSeek-OCR: Contexts Optical Compression

Haoran Wei, Yaofeng Sun, Yukun Li

DeepSeek- AI

Abstract

We present DeepSeek- OCR as an initial investigation into the feasibility of compressing long contexts via optical 2D mapping. DeepSeek- OCR consists of two components: DeepEncoder and DeepSeek3B- MoE- A570M as the decoder. Specifically, DeepEncoder serves as the core engine, designed to maintain low activations under high- resolution input while achieving high compression ratios to ensure an optimal and manageable number of vision tokens. Experiments show that when the number of text tokens is within 10 times that of vision tokens (i.e., a compression ratio (< 10\times) ), the model can achieve decoding (OCR) precision of (97\%) . Even at a compression ratio of (20\times) , the OCR accuracy still remains at about (60\%) . This shows considerable promise for research areas such as historical long- context compression and memory forgetting mechanisms in LLMs. Beyond this, DeepSeek- OCR also demonstrates high practical value. On OmniDocBench, it surpasses GOT- OCR2.0 (256 tokens/page) using only 100 vision tokens, and outperforms MinerU2.0 ( (6000+) tokens per page on average) while utilizing fewer than 800 vision tokens. In production, DeepSeek- OCR can generate training data for LLMs/VLMs at a scale of (200k+) pages per day (a single A100- 40G). Codes and model weights are publicly accessible at http://github.com/deepseek– ai/DeepSeek- OCR.

Figure 1 Description:

(a) Compression on Fox benchmark

This bar chart shows the compression of various encoders on the Fox benchmark. The x-axis represents the text tokens in per page (ground truth), and the y-axis represents the overall precision (%). The different colors represent different encoders, as indicated in the legend. The chart shows that the overall precision of the encoders decreases as the text tokens increase. This is because the encoders are designed to compress text, and as the text tokens increase, the encoders have less to compress, resulting in lower overall precision.

(b) Performance on Omnidocbench

This scatter plot shows the performance of various deep learning models on the Omnidocbench benchmark. The x-axis represents the average vision tokens per image, and the y-axis represents the overall precision (%). The different colors represent different deep learning models, as indicated in the legend. The chart shows that the overall precision of the models decreases as the average vision tokens per image increases. This is because the models are designed to compress images, and as the average vision tokens per image increases, the models have less to compress, resulting in lower overall precision.

(a) Compression on Fox benchmark

This bar chart shows the compression of various encoders on the Fox benchmark. The x-axis represents the text tokens in per page (ground truth), and the y-axis represents the overall precision (%). The different colors represent different encoders, as indicated in the legend. The chart shows that the overall precision of the encoders decreases as the text tokens increase. This is because the encoders are designed to compress text, and as the text tokens increase, the encoders have less to compress, resulting in lower overall precision.

(b) Performance on Omnidocbench

This scatter plot shows the performance of various deep learning models on the Omnidocbench benchmark. The x-axis represents the average vision tokens per image, and the y-axis represents the overall precision (%). The different colors represent different deep learning models, as indicated in the legend. The chart shows that the overall precision of the models decreases as the average vision tokens per image increases. This is because the models are designed to compress images, and as the average vision tokens per image increases, the models have less to compress, resulting in lower overall precision.

Figure 1 | Figure (a) shows the compression ratio (number of text tokens in ground truth/number of vision tokens model used) testing on Fox [21] benchmark; Figure (b) shows performance comparisons on OmniDocBench [27]. DeepSeek-OCR can achieve state-of-the-art performance among end-to-end models enjoying the fewest vision tokens.

Contents

1 Introduction 3

2 Related Works 4

2.1 Typical Vision Encoders in VLMs 4

2.2 End-to-end OCR Models 4

3 Methodology 5

3.1 Architecture 5

3.2 DeepEncoder 5

3.2.1 Architecture of DeepEncoder 5

3.2.2 Multiple resolution support 6

3.3 The MoE Decoder 7

3.4 Data Engine 7

3.4.1 OCR 1.0 data 7

3.4.2 OCR 2.0 data 8

3.4.3 General vision data 9

3.4.4 Text-only data 9

3.5 Training Pipelines 9

3.5.1 Training DeepEncoder 10

3.5.2 Training DeepSeek- OCR 10

4 Evaluation 10

4.1 Vision- text Compression Study 10

4.2 OCR Practical Performance 12

4.3 Qualitative Study 12

4.3.1 Deep parsing 12

4.3.2 Multilingual recognition 16

4.3.3 General vision understanding 17

5 Discussion 18

6 Conclusion 19

1. Introduction

Current Large Language Models (LLMs) face significant computational challenges when processing long textual content due to quadratic scaling with sequence length. We explore a potential solution: leveraging visual modality as an efficient compression medium for textual information. A single image containing document text can represent rich information using substantially fewer tokens than the equivalent digital text, suggesting that optical compression through vision tokens could achieve much higher compression ratios.

This insight motivates us to reexamine vision- language models (VLMs) from an LLM- centric perspective, focusing on how vision encoders can enhance LLMs’ efficiency in processing textual information rather than basic VQA [12, 16, 24, 32, 41] what humans excel at. OCR tasks, as an intermediate modality bridging vision and language, provide an ideal testbed for this vision- text compression paradigm, as they establish a natural compression- decompression mapping between visual and textual representations while offering quantitative evaluation metrics.

Accordingly, we present DeepSeek- OCR, a VLM designed as a preliminary proof- of- concept for efficient vision- text compression. Our work makes three primary contributions:

First, we provide comprehensive quantitative analysis of vision- text token compression ratios. Our method achieves (96\% +) OCR decoding precision at (9 – 10\times) text compression, (\sim 90\%) at (10 – 12\times) compression, and (\sim 60\%) at (20\times) compression on Fox [21] benchmarks featuring diverse document layouts (with actual accuracy being even higher when accounting for formatting differences between output and ground truth), as shown in Figure 1(a). The results demonstrate that compact language models can effectively learn to decode compressed visual representations, suggesting that larger LLMs could readily acquire similar capabilities through appropriate pretraining design.

Second, we introduce DeepEncoder, a novel architecture that maintains low activation memory and minimal vision tokens even with high- resolution inputs. It serially connects window attention and global attention encoder components through a (16\times) convolutional compressor. This design ensures that the window attention component processes a large number of vision tokens, while the compressor reduces vision tokens before they enter the dense global attention component, achieving effective memory and token compression.

Third, we develop DeepSeek- OCR based on DeepEncoder and DeepSeek3B- MoE [19, 20]. As shown in Figure 1(b), it achieves state- of- the- art performance within end- to- end models on OmniDocBench while using the fewest vision tokens. Additionally, we equip the model with capabilities for parsing charts, chemical formulas, simple geometric figures, and natural images to enhance its practical utility further. In production, DeepSeek- OCR can generate 33 million pages of data per day for LLMs or VLMs using 20 nodes (each with 8 A100- 40G GPUs).

In summary, this work presents a preliminary exploration of using visual modality as an efficient compression medium for textual information processing in LLMs. Through DeepSeek- OCR, we demonstrate that vision- text compression can achieve significant token reduction (7- 20x) for different historical context stages, offering a promising direction for addressing long- context challenges in large language models. Our quantitative analysis provides empirical guidelines for VLM token allocation optimization, while the proposed DeepEncoder architecture showcases practical feasibility with real- world deployment capabilities. Although focused on OCR as a proof- of- concept, this paradigm opens new possibilities for rethinking how vision and language modalities can be synergistically combined to enhance computational efficiency in large- scale text processing and agent systems.

Figure 1 Description: The figure illustrates the architecture of the VIT (Vision Transformer) model, which is a type of deep learning model used for various computer vision tasks. The architecture is depicted in a step-by-step manner, showing the flow of data through different layers and components of the model.

1. Input Data:

   • The input data is represented as a 1024-dimensional vector, which is then processed through two parallel VIT blocks.

2. VIT Blocks:

   • The first VIT block consists of a VIT module followed by a down-sample operation. This process reduces the dimensionality of the input data.
   • The second VIT block also includes a VIT module followed by a down-sample operation, further reducing the dimensionality.

3. LLM (Large Language Model):

   • After the VIT blocks, the data is passed through an LLM, which processes the information and generates the final output.

4. Output Generation:

   • The output from the LLM is then used to generate a sequence of tokens. The tokens are represented as (w//14(16))×(h//14(16)), where w and h are the width and height of the output, and (16) represents the number of tokens in each dimension.

5. Token Generation:

   • The tokens are then used to generate the final output, which is a sequence of words or tokens.

6. Loss Functions:

   • The model uses multiple loss functions to optimize the performance of the VIT model. These loss functions include:
     • Unsupported Pipeline Parallelism: Ensures that the pipeline operations are executed correctly.
     • Unsupported Extreme Resolution: Ensures that the model can handle images with varying resolutions.
     • Two Pre-processes: Likely refers to pre-processing steps such as normalization or data augmentation.
     • Hard to Deployment: Indicates that the model may face challenges in deployment due to its complexity.
     • Low Native Resolution: Ensures that the model can handle images at their native resolution.
     • Overly Small Patches: Refers to the size of the input patches being processed.
     • Small Global View: Indicates that the model may not capture global context effectively.
     • Large Activations: Refers to the number of weights in the model’s layers.
     • Slow Inference Speed: Indicates that the model may have slower performance compared to other models.

The figure also includes a note on the quantization process, where the weights and activations are quantized to reduce the model’s size and improve efficiency. The quantization process is applied to both the VIT modules and the LLM, ensuring that the model remains efficient even with reduced precision.

Overall, the figure provides a detailed overview of the VIT model architecture, highlighting its key components, the flow of data through the model, and the various loss functions used to optimize its performance.

Figure 2 | Typical vision encoders in popular VLMs. Here are three types of encoders commonly used in current open-source VLMs, all of which suffer from their respective deficiencies.

2. Related Works

2.1. Typical Vision Encoders in VLMs

Current open- source VLMs employ three main types of vision encoders, as illustrated in Figure 2. The first type is a dual- tower architecture represented by Vary [36], which utilizes parallel SAM [17] encoder to increase visual vocabulary parameters for high- resolution image processing. While offering controllable parameters and activation memory, this approach suffers from significant drawbacks: it requires dual image preprocessing that complicates deployment and makes encoder pipeline parallelism challenging during training. The second type is tile- based method exemplified by InternVL2.0 [8], which processes images by dividing them into small tiles for parallel computation, reducing activation memory under high- resolution settings. Although capable of handling extremely high resolutions, this approach has notable limitations due to its typically low native encoder resolution (below (512 \times 512) ), causing large images to be excessively fragmented and resulting in numerous vision tokens. The third type is adaptive resolution encoding represented by Qwen2- VL [35], which adopts the NaViT [10] paradigm to directly process full images through patch- based segmentation without tile parallelization. While this encoder can handle diverse resolutions flexibly, it faces substantial challenges with large images due to massive activation memory consumption that can cause GPU memory overflow, and sequence packing requires extremely long sequence lengths during training. Long vision tokens will slow down both prefill and generation phases of inference.

2.2. End-to-end OCR Models

OCR, particularly document parsing task, has been a highly active topic in the image- to- text domain. With the advancement of VLMs, a large number of end- to- end OCR models have emerged, fundamentally transforming the traditional pipeline architecture (which required separate detection and recognition expert models) by simplifying OCR systems. Nougat [6] first employs end- to- end framework for academic paper OCR on arXiv, demonstrating the potential of models in handling dense perception tasks. GOT- OCR2.0 [38] expands the scope of OCR2.0 to include more synthetic image parsing tasks and designs an OCR model with performance- efficiency trade- offs, further highlighting the potential of end- to- end OCR researches. Additionally, general vision models such as Qwen- VL series [35], InternVL series [8], and many their derivatives continuously enhance their document OCR capabilities to explore dense visual perception boundaries. However, a crucial research question that current models have not addressed is: for a document containing 1000 words, how many vision tokens are at least needed for decoding? This question holds significant importance for research in the principle that “a picture is worth a thousand words.”

Figure 1 Description: The image is a diagram illustrating the architecture of a deep learning model used for text-to-image generation. The model consists of several components: an input image, a local attention mechanism, a tokenizer, a CLIP VIT 300M model, a deep encoder, a decoder, and an output.

1. Input: The process begins with an input image, which is a document containing text. The image is fed into the system.

2. Local Attention: The input image is processed by a local attention mechanism. This mechanism highlights important regions of the image that are relevant for generating the image. The output of this step is a set of patches, each with a size of 16×16.

3. Tokenizer: The patches are then tokenized, which involves converting the patches into tokens. This step is crucial for preparing the data for further processing.

4. Local Attention: The tokens are processed by another local attention mechanism. This step refines the tokens by focusing on the most relevant parts of the image.

5. Conv: The refined tokens are passed through a convolutional layer. This layer helps in capturing spatial features from the tokens.

6. Vision Tokens: The output from the convolutional layer is then used to generate vision tokens. These tokens represent the visual information from the input image.

7. Deep Encoder: The vision tokens are fed into a deep encoder. The encoder processes these tokens to generate a global attention token. This token helps in capturing the global context of the image.

8. CLIP VIT 300M: The global attention token is then passed through a CLIP VIT 300M model. This model processes the global context to generate a prompt. The prompt is a textual description of the image that the model will use to generate the image.

9. Decoder: The prompt is fed into the decoder, which generates the final output image. The decoder uses the global attention token and the prompt to produce the image.

10. Output: The final output is the generated image, which is a faithful reproduction of the input image based on the prompt.

The diagram also shows the flow of data and the interactions between different components of the model. The process starts with the input image, goes through local attention and tokenization, followed by convolutional and deep encoding layers, and finally results in the generation of the output image.

Figure 3 | The architecture of DeepSeek-OCR. DeepSeek-OCR consists of a DeepEncoder and a DeepSeek-3B-MoE decoder. DeepEncoder is the core of DeepSeek-OCR, comprising three components: a SAM [17] for perception dominated by window attention, a CLIP [29] for knowledge with dense global attention, and a (16\times) token compressor that bridges between them.

3. Methodology

3.1. Architecture

As shown in Figure 3, DeepSeek- OCR enjoys a unified end- to- end VLM architecture consisting of an encoder and a decoder. The encoder (namely DeepEncoder) is responsible for extracting image features and tokenizing as well as compressing visual representations. The decoder is used for generating the required result based on image tokens and prompts. DeepEncoder is approximately 380M in parameters, mainly composed of an 80M SAM- base [17] and a 300M CLIP- large [29] connected in series. The decoder adopts a 3B MoE [19, 20] architecture with 570M activated parameters. In the following paragraphs, we will delve into the model components, data engineering, and training skills.

3.2. DeepEncoder

To explore the feasibility of contexts optical compression, we need a vision encoder with the following features: 1. Capable of processing high resolutions; 2. Low activation at high resolutions; 3. Few vision tokens; 4. Support for multiple resolution inputs; 5. Moderate parameter count. However, as described in the Section 2.1, current open- source encoders cannot fully satisfy all these conditions. Therefore, we design a novel vision encoder ourselves, named DeepEncoder.

3.2.1. Architecture of DeepEncoder

DeepEncoder mainly consists of two components: a visual perception feature extraction component dominated by window attention, and a visual knowledge feature extraction component with dense global attention. To benefit from the pretraining gains of previous works, we use SAM- base (patch- size 16) and CLIP- large as the main architectures for the two components respectively. For CLIP, we remove the first patch embedding layer since its input is no longer images but output tokens from the previous pipeline. Between the two components, we borrow from Vary [36] and use a 2- layer convolutional module to perform (16\times) downsampling of vision tokens. Each convolutional layer has a kernel size of 3, stride of 2, padding of 1, and channels increase from 256 to 1024. Assuming we input a (1024\times 1024) image, the DeepEncoder will segment it into (1024 / 16\times 1024 / 16 = 4096) patch tokens. Since the first half of encoder is dominated by window attention and only 80M, the activation is acceptable. Before entering global attention,

====================================================================== PAGE 6

Figure 1 Description: The image displays three different types of electronic devices, each with a unique design and specifications. On the left, there is a small, rectangular device labeled “W:512|640” with a screen displaying a colorful interface. The device has a black casing with a visible antenna on the top right corner. Below the screen, there are two buttons, one labeled “RESIZE” and the other “TONE.”

In the center, there is a larger device with a similar design, but it has a black casing with a visible antenna on the top right corner. The screen displays a blue interface with a circular dial in the center, surrounded by various icons and text. Below the screen, there are two buttons, one labeled “TONE” and the other “TONE.”

On the right, there is a device with a black casing and a visible antenna on the top right corner. The screen displays a blue interface with a circular dial in the center, surrounded by various icons and text. Below the screen, there are two buttons, one labeled “TONE” and the other “TONE.”

Each device has a label at the bottom with technical specifications such as “Mode: Tiny|Small,” “Token: 64|100,” and “R:1-(H-W)/W.” The background of the image is white, and the text is primarily in black with some blue highlights. The layout is organized and symmetrical, with each device occupying its own section of the image.

Figure 4 | To test model performance under different compression ratios (requiring different numbers of vision tokens) and enhance the practicality of DeepSeek-OCR, we configure it with multiple resolution modes.

the 4096 tokens go through the compression module and the token count becomes (4096 / 16 = 256) , thus making the overall activation memory controllable.

Table 1 | Multi resolution support of DeepEncoder. For both research and application purposes, we design DeepEncoder with diverse native resolution and dynamic resolution modes.

Mode	Native Resolution			Dynamic Resolution
	Tiny	Small	Base	Larger	Gundam
Resolution	512	640	1024	1280	640+1024
Tokens	64	100	256	400	n×100+256
Process	resize	resize	padding	padding	resize + padding

3.2.2. Multiple resolution support

Suppose we have an image with 1000 optical characters and we want to test how many vision tokens are needed for decoding. This requires the model to support a variable number of vision tokens. That is to say the DeepEncoder needs to support multiple resolutions.

We meet the requirement aforementioned through dynamic interpolation of positional encodings, and design several resolution modes for simultaneous model training to achieve the capability of a single DeepSeek- OCR model supporting multiple resolutions. As shown in Figure 4, DeepEncoder mainly supports two major input modes: native resolution and dynamic resolution. Each of them contains multiple sub- modes.

Native resolution supports four sub- modes: Tiny, Small, Base, and Large, with corresponding resolutions and token counts of (512 \times 512) (64), (640 \times 640) (100), (1024 \times 1024) (256), and (1280 \times 1280) (400) respectively. Since Tiny and Small modes have relatively small resolutions, to avoid wasting vision tokens, images are processed by directly resizing the original shape. For Base and Large modes, in order to preserve the original image aspect ratio, images are padded to the corresponding size. After padding, the number of valid vision tokens is less than the actual number of vision tokens, with the calculation formula being:

[N{valid} = \lceil N{actual}\times [1 – ((max(w,h) – min(w,h)) / (max(w,h)))]\rceil \quad (1)]

where (w) and (h) represent the width and height of the original input image.

====================================================================== PAGE 7

Dynamic resolution can be composed of two native resolutions. For example, Gundam mode consists of (n \times 640 \times 640) tiles (local views) and a (1024 \times 1024) global view. The tiling method following InternVL2.0 [8]. Supporting dynamic resolution is mainly for application considerations, especially for ultra- high- resolution inputs (such as newspaper images). Tiling is a form of secondary window attention that can effectively reduce activation memory further. It’s worth noting that due to our relatively large native resolutions, images won’t be fragmented too much under dynamic resolution (the number of tiles is controlled within the range of 2 to 9). The vision token number output by the DeepEncoder under Gundam mode is: (n \times 100 + 256) , where (n) is the number of tiles. For images with both width and height smaller than 640, (n) is set to 0, i.e., Gundam mode will degrade to Base mode.

Gundam mode is trained together with the four native resolution modes to achieve the goal of one model supporting multiple resolutions. Note that Gundam- master mode ( (1024 \times 1024) local views (+ 1280 \times 1280) global view) is obtained through continued training on a trained DeepSeek- OCR model. This is mainly for load balancing, as Gundam- master’s resolution is too large and training it together would slow down the overall training speed.

3.3. The MoE Decoder

Our decoder uses the DeepSeekMoE [19, 20], specifically DeepSeek- 3B- MoE. During inference, the model activates 6 out of 64 routed experts and 2 shared experts, with about 570M activated parameters. The 3B DeepSeekMoE is very suitable for domain- centric (OCR for us) VLM research, as it obtains the expressive capability of a 3B model while enjoying the inference efficiency of a 500M small model.

The decoder reconstructs the original text representation from the compressed latent vision tokens of DeepEncoder as:

[f{\mathrm{dec}}:\mathbb{R}^{n\times d{\mathrm{latent}}}\to \mathbb{R}^{N\times d{\mathrm{text}}}; \quad \hat{\mathbf{X}} = f{\mathrm{dec}}(\mathbf{Z})\quad \mathrm{where} n\leq N \quad (2)]

where (\mathbf{Z} \in \mathbb{R}^{n \times d{\mathrm{latent}}}) are the compressed latent(vision) tokens from DeepEncoder and (\hat{\mathbf{X}} \in \mathbb{R}^{N \times d{\mathrm{text}}}) is the reconstructed text representation. The function (f_{\mathrm{dec}}) represents a non- linear mapping that can be effectively learned by compact language models through OCR- style training. It is reasonable to conjecture that LLMs, through specialized pretraining optimization, would demonstrate more natural integration of such capabilities.

3.4. Data Engine

We construct complex and diverse training data for DeepSeek- OCR, including OCR 1.0 data, which mainly consists of traditional OCR tasks such as scene image OCR and document OCR; OCR 2.0 data, which mainly includes parsing tasks for complex artificial images, such as common charts, chemical formulas, and plane geometry parsing data; General vision data, which is mainly used to inject certain general image understanding capabilities into DeepSeek- OCR and preserve the general vision interface.

3.4.1. OCR 1.0 data

Document data is the top priority for DeepSeek- OCR. We collect 30M pages of diverse PDF data covering about 100 languages from the Internet, with Chinese and English accounting for approximately 25M and other languages accounting for 5M. For this data, we create two types of ground truth: coarse annotations and fine annotations. Coarse annotations are extracted

====================================================================== PAGE 8

Figure 1 Description: The image displays a page of a mathematics textbook or worksheet. The page is titled “Ground truth image” and is labeled as “2. nácín:”. It contains a diagram of a triangle with vertices labeled ( PE ), ( CE ), ( SU ), ( NE ), ( PO ), and ( UT ). The triangle is divided into two parts by a line labeled ( SR ). The diagram is accompanied by a mathematical problem involving the area of a triangle and the lengths of its sides.

Below the diagram, there is a mathematical problem statement that reads: “14. Maša želi popuniti tablicu tako da u svaku ćeliju upiše jedan broj. Za sada je upisala dva broja kako je prikazano na slici. Tablicu želi popuniti tako da je broj svih upisanih brojeva 35, broj brojeva u prve tri ćelije je 22, a broj brojeva u posljednje tri ćelije je 25. Koliki je umnožak brojeva koje će upisati u sive ćelije?”

The problem is followed by a series of multiple-choice answers labeled A) 63, B) 108, C) 0, D) 48, and E) 39, with the correct answer being 63.

The page also includes a section with instructions and guidelines for the problem, including the use of a formula for the area of a triangle and the calculation of the area of the triangle based on the given lengths of its sides. The text is in Czech, and the page is numbered 2 at the top right corner.

Figure 5 | OCR 1.0 fine annotations display. We format the ground truth into an interleaved layout and text format, where each paragraph of text is preceded by the coordinates and label of it in the original image. All coordinates are normalized into 1000 bins.

directly from the full dataset using fitz, aimed at teaching the model to recognize optical text, especially in minority languages. Fine annotations include 2M pages each for Chinese and English, labeled using advanced layout models (such as PP- DocLayout [33]) and OCR models (such as MinuerU [34] and GOT- OCR2.0 [38]) to construct detection and recognition interleaved data. For minority languages, in the detection part, we find that the layout model enjoys certain generalization capabilities. In the recognition part, we use fitz to create small patch data to train a GOT- OCR2.0, then use the trained model to label small patches after layout processing, employing a model flywheel to create 600K data samples. During the training of DeepSeek- OCR, coarse labels and fine labels are distinguished using different prompts. The ground truth for fine annotation image- text pairs can be seen in Figure 5. We also collect 3M Word data, constructing high- quality image- text pairs without layout by directly extracting content. This data mainly brings benefits to formulas and HTML- formatted tables. Additionally, we select some open- source data [28, 37] as supplements.

For natural scene OCR, our model mainly supports Chinese and English. The image data sources come from LAION [31] and Wukong [13], labeled using PaddleOCR [9], with 10M data samples each for Chinese and English. Like document OCR, natural scene OCR can also control whether to output detection boxes through prompts.

3.4.2. OCR 2.0 data

Following GOT- OCR2.0 [38], we refer to chart, chemical formula, and plane geometry parsing data as OCR 2.0 data. For chart data, following OneChart [7], we use pyecharts and matplotlib

====================================================================== PAGE 9

Figure 1 Description:

(a) Image-text ground truth of chart

This section contains a bar chart and a pie chart. The bar chart shows the distribution of different types of data points across various categories, while the pie chart illustrates the proportion of each category.

(b) Image-text ground truth of geometry

This section contains a 3D geometric diagram with various labeled points and lines. The points are connected by lines, forming a network that represents the geometric relationships between them. The diagram includes labels for different parts of the network, such as “E,” “S,” “R,” and “X,” along with numerical values indicating specific measurements or distances.

The image also includes a table with numerical data, which appears to be related to the geometric network. The table contains rows and columns with numerical values, possibly representing measurements or ratios between different parts of the network.

Overall, the figure provides a detailed comparison between image-text ground truth and geometric ground truth, highlighting the differences and similarities in how these two types of data are structured and represented.

Figure 6 | For charts, we do not use OneChart’s [7] dictionary format, but instead use HTML table format as labels, which can save a certain amount of tokens. For plane geometry, we convert the ground truth to dictionary format, where the dictionary contains keys such as line segments, endpoint coordinates, line segment types, etc., for better readability. Each line segment is encoded using the Slow Perception [39] manner.

to render 10M images, mainly including commonly used line, bar, pie, and composite charts. We define chart parsing as image- to- HTML- table conversion task, as shown in Figure 6(a). For chemical formulas, we utilize SMILES format from PubChem as the data source and render them into images using RDKit, constructing 5M image- text pairs. For plane geometry images, we follow Slow Perception [39] for generation. Specifically, we use perception- ruler size as 4 to model each line segment. To increase the diversity of rendered data, we introduce geometric translation- invariant data augmentation, where the same geometric image is translated in the original image, corresponding to the same ground truth drawn at the centered position in the coordinate system. Based on this, we construct a total of 1M plane geometry parsing data, as illustrated in Figure 6(b).

3.4.3. General vision data

DeepEncoder can benefit from CLIP’s pretraining gains and has sufficient parameters to incorporate general visual knowledge. Therefore, we also prepare some corresponding data for DeepSeek- OCR. Following DeepSeek- VL2 [40], we generate relevant data for tasks such as caption, detection, and grounding. Note that DeepSeek- OCR is not a general VLM model, and this portion of data accounts for only (20\%) of the total data. We introduce such type of data mainly to preserve the general vision interface, so that researchers interested in our model and general vision task can conveniently advance their work in the future.

3.4.4. Text-only data

To ensure the model’s language capabilities, we introduced (10\%) of in- house text- only pretrain data, with all data processed to a length of 8192 tokens, which is also the sequence length for DeepSeek- OCR. In summary, when training DeepSeek- OCR, OCR data accounts for (70\%) , general vision data accounts for (20\%) , and text- only data accounts for (10\%) .

3.5. Training Pipelines

Our training pipeline is very simple and consists mainly of two stages: a). Training DeepEncoder independently; b). Training the DeepSeek- OCR. Note that the Gundam- master mode is obtained by continuing training on a pre- trained DeepSeek- OCR model with 6M sampled data. Since the training protocol is identical to other modes, we omit the detailed description hereafter.

====================================================================== PAGE 10

3.5.1. Training DeepEncoder

3.5.1. Training DeepEncoderFollowing Vary [36], we utilize a compact language model [15] and use the next token prediction framework to train DeepEncoder. In this stage, we use all OCR 1.0 and 2.0 data aforementioned, as well as 100M general data sampled from the LAION [31] dataset. All data is trained for 2 epochs with a batch size of 1280, using the AdamW [23] optimizer with cosine annealing scheduler [22] and a learning rate of 5e- 5. The training sequence length is 4096.

3.5.2. Training DeepSeek-OCR

After DeepEncoder is ready, we use data mentioned in Section 3.4 to train the DeepSeek- OCR. with the entire training process conducted on the HAI- LLM [14] platform. The entire model uses pipeline parallelism (PP) and is divided into 4 parts, with DeepEncoder taking two parts and the decoder taking two parts. For DeepEncoder, we treat SAM and the compressor as the vision tokenizer, place them in PP0 and freeze their parameters, while treating the CLIP part as input embedding layer and place it in PP1 with unfrozen weights for training. For the language model part, since DeepSeek3B- MoE has 12 layers, we place 6 layers each on PP2 and PP3. We use 20 nodes (each with 8 A100- 40G GPUs) for training, with a data parallelism (DP) of 40 and a global batch size of 640. We use the AdamW optimizer with a step- based scheduler and an initial learning rate of 3e- 5. For text- only data, the training speed is 90B tokens/day, while for multimodal data, the training speed is 70B tokens/day.

Table 2 | We test DeepSeek- OCR’s vision- text compression ratio using all English documents with 600- 1300 tokens from the Fox [21] benchmarks. Text tokens represent the number of tokens after tokenizing the ground truth text using DeepSeek- OCR’s tokenizer. Vision Tokens=64 or 100 respectively represent the number of vision tokens output by DeepEncoder after resizing input images to 512×512 and 640×640.

Text Tokens	Vision Tokens =64		Vision Tokens=100
	Precision	Compression	Precision	Compression
600-700	96.5%	10.5×	98.5%	6.7×
700-800	93.8%	11.8×	97.3%	7.5×
800-900	83.8%	13.2×	96.8%	8.5×
900-1000	85.9%	15.1×	96.8%	9.7×
1000-1100	79.3%	16.5×	91.5%	10.6×
1100-1200	76.4%	17.7×	89.8%	11.3×
1200-1300	59.1%	19.7×	87.1%	12.6×

4. Evaluation

4.1. Vision-text Compression Study

4. Evaluation4.1. Vision- text Compression StudyWe select Fox [21] benchmarks to verify DeepSeek- OCR’s compression- decompression capability for text- rich documents, in order to preliminarily explore the feasibility and boundaries of contexts optical compression. We use the English document portion of Fox, tokenize the ground truth text with DeepSeek- OCR’s tokenizer (vocabulary size of approximately 129k), and select documents with 600- 1300 tokens for testing, which happens to be 100 pages. Since the number of text tokens is not large, we only need to test performance in Tiny and Small modes, where Tiny mode corresponds to 64 tokens and Small mode corresponds to 100 tokens. We use the prompt

====================================================================== PAGE 11

Table 3 | We use OmniDocBench [27] to test the performance of DeepSeek-OCR on real document parsing tasks. All metrics in the table are edit distances, where smaller values indicate better performance. “Tokens” represents the average number of vision tokens used per page, and (\mathrm{^{i + 200dpi}}) means using fitz to interpolate the original image to 200dpi. For the DeepSeek-OCR model, the values in parentheses in the “Tokens” column represent valid vision tokens, calculated according to Equation 1.

Model	Tokens	English				Chinese
		overall	text	formula	table	order	overall	text	formula
Pipline Models
Dolphin [11]	–	0.356	0.352	0.465	0.258	0.35	0.44	0.44	0.604
Marker [1]	–	0.296	0.085	0.374	0.609	0.116	0.497	0.293	0.688
Mathpix [2]	–	0.191	0.105	0.306	0.243	0.108	0.364	0.381	0.454
MinerU-2.1.1 [34]	–	0.162	0.072	0.313	0.166	0.097	0.244	0.111	0.581
MonkeyOCR-1.2B [18]	–	0.154	0.062	0.295	0.164	0.094	0.263	0.179	0.464
PPstructure-v3 [9]	–	0.152	0.073	0.295	0.162	0.077	0.223	0.136	0.535
End-to-end Models
Nougat [6]	2352	0.452	0.365	0.488	0.572	0.382	0.973	0.998	0.941
SmolDocking [25]	392	0.493	0.262	0.753	0.729	0.227	0.816	0.838	0.997
InternVL2-76B [8]	6790	0.44	0.353	0.543	0.547	0.317	0.443	0.29	0.701
Qwen2.5-VL-7B [5]	3949	0.316	0.151	0.376	0.598	0.138	0.399	0.243	0.5
OLMOCR [28]	3949	0.326	0.097	0.455	0.608	0.145	0.469	0.293	0.655
GOT-OCR2.0 [38]	256	0.287	0.189	0.360	0.459	0.141	0.411	0.315	0.528
OCRFlux-3B [3]	3949	0.238	0.112	0.447	0.269	0.126	0.349	0.256	0.716
GPT4o [26]	–	0.233	0.144	0.425	0.234	0.128	0.399	0.409	0.606
InternVL3-78B [42]	6790	0.218	0.117	0.38	0.279	0.095	0.296	0.21	0.533
Qwen2.5-VL-72B [5]	3949	0.214	0.092	0.315	0.341	0.106	0.261	0.18	0.434
dots.ocr [30]	3949	0.182	0.137	0.320	0.166	0.182	0.261	0.229	0.468
Gemini2.5-Pro [4]	–	0.148	0.055	0.356	0.13	0.049	0.212	0.168	0.439
MinerU2.0 [34]	6790	0.133	0.045	0.273	0.15	0.066	0.238	0.115	0.506
dots.ocr+200dpi [30]	5545	0.125	0.032	0.329	0.099	0.04	0.16	0.066	0.416
DeepSeek-OCR (end2end)
Tiny	64	0.386	0.373	0.469	0.422	0.283	0.361	0.307	0.635
Small	100	0.221	0.142	0.373	0.242	0.125	0.284	0.24	0.53
Base	256(182)	0.137	0.054	0.267	0.163	0.064	0.24	0.205	0.474
Large	400(285)	0.138	0.054	0.277	0.152	0.067	0.208	0.143	0.461
Gundam	795	0.127	0.043	0.269	0.134	0.062	0.181	0.097	0.432
Gundam-M+200dpi	1853	0.123	0.049	0.242	0.147	0.056	0.157	0.087	0.377

without layout: “<image>\nFree OCR.” to control the model’s output format. Nevertheless, the output format still cannot completely match Fox benchmarks, so the actual performance would be somewhat higher than the test results.

As shown in Table 2, within a (10\times) compression ratio, the model’s decoding precision can reach approximately (97\%) , which is a very promising result. In the future, it may be possible to achieve nearly (10\times) lossless contexts compression through text- to- image approaches. When the compression ratio exceeds (10\times) , performance begins to decline, which may have two reasons: one is that the layout of long documents becomes more complex, and another reason may be that long texts become blurred at (512\times 512) or (640\times 640) resolution. The first issue can be solved by rendering texts onto a single layout page, while we believe the second issue will become

====================================================================== PAGE 12

a feature of the forgetting mechanism. When compressing tokens by nearly (20x) , we find that precision can still approach (60\%) . These results indicate that optical contexts compression is a very promising and worthwhile research direction, and this approach does not bring any overhead because it can leverage VLM infrastructure, as multimodal systems inherently require an additional vision encoder.

Table 4 | Edit distances for different categories of documents in OmniDocBench. The results show that some types of documents can achieve good performance with just 64 or 100 vision tokens, while others require Gundam mode.

Type Mode	Book Slides	Financial Report	Textbook	Exam Paper	Magazine	Academic Papers	Notes	Newspaper Overall
Tiny	0.147 0.116	0.207	0.173	0.294	0.201	0.395	0.297	0.94
Small	0.085 0.111	0.079	0.147	0.171	0.107	0.131	0.187	0.744
Base	0.037 0.08	0.027	0.1	0.13	0.073	0.052	0.176	0.645
Large	0.038 0.108	0.022	0.084	0.109	0.06	0.053	0.155	0.353
Gundam	0.035 0.085	0.289	0.095	0.094	0.059	0.039	0.153	0.122
Guandam-M	0.052 0.09	0.034	0.091	0.079	0.079	0.048	0.1	0.099

4.2. OCR Practical Performance

DeepSeek-OCR is not only an experimental model; it has strong practical capabilities and can construct data for LLM/VLM pretraining. To quantify OCR performance, we test DeepSeek-OCR on OmniDocBench [27], with results shown in Table 3. Requiring only 100 vision tokens (640×640 resolution), DeepSeek-OCR surpasses GOT-OCR2.0 [38] which uses 256 tokens; with 400 tokens (285 valid tokens, 1280×1280 resolution), it achieves on-par performance with state-of-the-arts on this benchmark. Using fewer than 800 tokens (Gundam mode), DeepSeek-OCR outperforms MinerU2.0 [34] which needs nearly 7,000 vision tokens. These results demonstrate that our DeepSeek-OCR model is powerful in practical applications, and because the higher tokens compression, it enjoys a higher research ceiling.

As shown in Table 4, some categories of documents require very few tokens to achieve satisfactory performance, such as slides which only need 64 vision tokens. For book and report documents, DeepSeek-OCR can achieve good performance with only 100 vision tokens.Combined with the analysis from Section 4.1, this may be because most text tokens in these document categories are within 1,000, meaning the vision-token compression ratio does not exceed 10x. For newspapers, Gundam or even Gundam-master mode is required to achieve acceptable edit distances, because the text tokens in newspapers are 4-5,000, far exceeding the 10x compression of other modes. These experimental results further demonstrate the boundaries of contexts optical compression, which may provide effective references for researches on the vision token optimization in VLMs and context compression, forgetting mechanisms in LLMs.

4.3. Qualitative Study

4.3.1. Deep parsing

DeepSeek-OCR possesses both layout and OCR 2.0 capabilities, enabling it to further parse images within documents through secondary model calls, a feature we refer to as “deep parsing”.As shown in Figures 7,8,9,10, our model can perform deep parsing on charts, geometry, chemical formulas, and even natural images, requiring only a unified prompt.

====================================================================== PAGE 13

Figure 1 Description: Top Section:

• Title: Macro news and views
• Subtitle: We provide a brief snapshot on the most important economies for the global markets

Left Section:

• Title: US
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the US effective tariff rate from 4.4pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m recession outlook to 30% from 15% to reflect our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the US.

Middle Section:

• Title: Japan
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the US effective tariff rate from 4.4pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the US.

Right Section:

• Title: Euro
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the Euro effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the Euro.

Bottom Section:

• Title: Germany
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the German effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the German.

Middle Section:

• Title: France
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the French effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the French.

Middle Section:

• Title: Italy
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the Italian effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the Italian.

Middle Section:

• Title: Spain
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the Spanish effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the Spanish.

Middle Section:

• Title: Euro area
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the Euro area effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the Euro area.

Middle Section:

• Title: Japan
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the Japanese effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the Japanese effective tariff rate.

Middle Section:

• Title: Germany
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the German effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the German effective tariff rate.

Middle Section:

• Title: France
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the French effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the French effective tariff rate.

Middle Section:

• Title: Italy
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the Italian effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7% from 2.4%, OAG20—our first below- consensus call in 2.5 years—and slightly raised our 2029 unemployment rate forecast to 4.2% from 4.1% and our 12m forecast.
  • We expect to see 2025 and 2029 rate hikes in the Italian effective tariff rate.

Middle Section:

• Title: Spain
• Subtitle: Latest GS proprietary datapoints/major changes in views
• Content:
  • We now assume a 10pc increase in the Spanish effective tariff rate from 4.5pp point as reciprocal tariffs and further increases in product-specific tariffs now seem likely.
  • We raised our 2025 GDP growth forecast to 1.7%

Figure 7 | In the field of financial research reports, the deep parsing mode of DeepSeek-OCR can be used to obtain structured results of charts within documents. Charts are a crucial form of data representation in finance and scientific fields, and the chart structured extraction is an indispensable capability for future OCR models.

====================================================================== PAGE 14

Figure 1 Description: The image is a photograph of a group of people sitting around a table in a conference room. There are 11 people in the room, 10 men and 1 woman. They are all wearing business suits. The table is covered in papers and folders. The people are all sitting in chairs, and they are all looking at the woman at the head of the table. The woman is smiling, and she is holding a pen. She is looking at the papers on the table. The people at the table are all listening attentively to the woman. The image is taken from a high angle, and it is clear that the room is large and spacious. The image is well-lit, and the colors are bright and vibrant. The style of the image is photojournalistic.

Input image

Figure 2 Description: A classroom of children sitting on the floor in front of a teacher. The teacher is sitting on a chair in front of the children. She is reading a book to them. The children are sitting on the floor in front of the teacher. They are all looking at the teacher. The children are of different ages. There are two boys and two girls. The boys are wearing blue shirts and the girls are wearing pink shirts. The teacher is wearing a white shirt and black pants. She has long brown hair. The children are all sitting in different positions. Some of them are sitting cross-legged, some are sitting on their knees, and some are sitting on their chairs. The teacher is holding the book in her hands. She is smiling at the children. The children are all looking at the book. They are all listening to the teacher. The classroom is a bright and airy place. There are windows on the walls. There is a rug on the floor. The teacher is standing in the middle of the room. She is looking at the children. She is smiling at them. The children are all looking at the teacher. They are all smiling. The classroom is a happy place. The children are all enjoying their time with the teacher. The image is a photograph. It is taken from the perspective of a child sitting in the classroom. The image is in focus and the colors are bright and vibrant. The image captures the joy and happiness of a classroom.

Result

Figure 3 Description: Left Side (Image):

• The image shows a classroom setting with children seated on the floor, engaged in an activity.
• The children are of various ages and are dressed in casual clothing.
• The classroom has colorful decorations on the walls, including a large tree mural and various educational posters.
• The floor is covered with a colorful rug, and there are several children sitting on it.
• The children appear to be focused on the activity at hand, with some looking at the teacher or materials.

Right Side (Storybook Reading for Young Dual Language Learners):

• The image shows a book titled “Storybook Reading for Young Dual Language Learners” by Cristina Gilanders and Dina C. Castro.
• The book cover features a colorful illustration of children engaged in a reading activity.
• The text on the book cover is in both English and Spanish, indicating bilingual content.
• The book is open to a page with text and illustrations, likely from the story being read aloud.

Connecting Elements:

• A dotted line connects the image of the classroom to the book cover, indicating that the book is being used as a visual aid or example for the activity described in the text.
• The text on the book cover reads: “In a community of practice meeting, teachers discuss their experiences reading aloud to dual language learners. Susan: When I am reading a story, the Latino children in my class just sit there. They look at me, but you can tell that they are not engaged in the story. Lisa: That happens in my class too. The little girls play with their hair, and the boys play with their shoes. Beverly: And when you ask questions about the story, children who speak English take over and you can’t get an answer from the Latino children. Facilitator: What do you think is happening here? Lisa: I think they just don’t understand what the story is about. Facilitator: How can we help them understand the story so they can participate? RESEARCHERS WIDELY RECOMMEND storybook reading for promoting the early language and literacy of young children. By listening to stories, children learn about written syntax and vocabulary and develop phonological awareness and concepts of print. all of which are closely linked to learning to read and write (National Early Literacy Panel 2008). Teachers usually know a read-aloud experience has been effective because they see the children maintain their interest in the story, relate different aspects of the story to their own experiences, describe the illustrations, and ask questions about the characters and plot. However, listening to a story read aloud can be a very different experience for children who speak a language other than English. What happens when the children are read to in a language they are just beginning to learn? What happens when an English-speaking teacher reads a story to a group of children who are learning English as a second language? As illustrated in the vignette at the beginning of this article, teachers often describe young dual language learners in their class as distracted and unengaged during read-aloud sessions in English. In this article, we describe teaching strategies that English-speaking teachers can use when reading aloud to young dual language learners. These strategies are part of the Nuestros Niños”

This figure illustrates the connection between the classroom activity and the use of a storybook to support dual language learners, emphasizing the importance of creating an engaging and inclusive learning environment.

Figure 8 | For books and articles, the deep parsing mode can output dense captions for natural images in the documents. With just a prompt, the model can automatically identify what type of image it is and output the required results.

====================================================================== PAGE 15

Figure 1 Description: The figure illustrates the process of converting a document into a structured format using the PCT/IB2013/05771 dataset. The document is analyzed to extract key information, which is then converted into a structured format suitable for further processing.

Detailed Description:

1. Document Content:

   • The document contains a chemical structure of a compound, specifically a pyrenyl group attached to a benzene ring.
   • The chemical structure is labeled with various parts such as functional groups, substituents, and aromatic rings.

2. Conversion to Structured Format:

   • The document is then converted into a structured format using the PCT/IB2013/05771 dataset.
   • The structured format includes fields such as WO 2013/171642, PCT/IB2013/05771, and WO 2013/171642.

3. Example 24:

   • The structured format is used to create an example entry.
   • The example entry includes the chemical structure, functional groups, and other relevant information.
   • The example entry is labeled as Example 24.

4. Input Image:

   • The input image is a chemical structure diagram.
   • The diagram shows the molecular structure of the compound, with different parts labeled and connected.

5. Result:

   • The result shows the converted structured format.
   • The result includes the converted chemical structure, along with additional information such as the WO 2013/171642 and PCT/IB2013/05771 fields.

Summary:

The figure demonstrates the conversion of a chemical document into a structured format using the PCT/IB2013/05771 dataset. This structured format includes detailed information about the chemical structure, including functional groups, substituents, and aromatic rings. The converted structured format is then used to create an example entry and the result, providing a clear example of how the document is processed and converted into a structured format suitable for further analysis or use in a database or system.

Figure 9 | DeepSeek-OCR in deep parsing mode can also recognize chemical formulas within chemical documents and convert them to SMILES format. In the future, OCR 1.0+2.0 technology may play a significant role in the development of VLM/LLM in STEM fields.

====================================================================== PAGE 16

Figure 1 Description: Top Section (Title Page):

• Title: 八年级数学下册几何证明题练习
  • Description:
    • 1. 已知：△ABC的两条高BD、CE交于点F，点M、N、分别是AF、BC的中点，连接ED、MN。
    • (1) 证明：MN垂直于BD。
    • (2) 若∠CBD=∠DCE=45°，判断以M、E、N、D为顶点的四边形是矩形，并证明你的结论。
    • 2. 四边形ABCD是正方形，△ABF是等腰直角三角形，∠ABF=90°，BE=EF，连接DF、G为DF的中点，连接EG、CG、EC。
    • (1) 证明：1. 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠ECG的度数；
    • (2) 证明：1. 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠ECG的度数；
    • (3) 证明：1. 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠ECG的度数；
    • (4) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (5) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (6) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (7) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (8) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (9) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (10) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (11) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (12) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (13) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (14) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (15) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (16) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (17) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (18) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (19) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (20) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (21) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (22) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (23) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (24) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (25) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (26) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (27) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (28) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (29) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (30) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (31) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (32) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (33) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (34) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (35) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (36) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (37) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (38) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (39) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (40) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (41) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (42) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (43) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (44) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (45) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (46) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (47) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (48) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (49) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (50) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (51) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (52) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (53) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (54) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (55) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (56) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (57) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (58) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (59) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (60) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (61) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (62) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (63) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及△EGC的度数；
    • (64) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及∠EGC的度数；
    • (65) 若点E在CB边的延长线上，直接写出EG与CG的位置关系及

Figure 10 | DeepSeek-OCR also possesses the capability to copy (structure) simple planar geometric figures. Due to the intricate interdependencies among line segments in geometric shapes, parsing geometry task is extremely challenging and has a long way to go.

4.3.2. Multilingual recognition

PDF data on the Internet contains not only Chinese and English, but also a large amount of multilingual data, which is also crucial when training LLMs. For PDF documents, DeepSeek- OCR can handle nearly 100 languages. Like Chinese and English documents, multilingual data also supports both layout and non- layout OCR formats. The visualization results are shown in Figure 11, where we select Arabic and Sinhala languages to demonstrate results.

====================================================================== PAGE 17

Figure 1 Description: NoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNone NoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNone NoneNone NoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNoneNone NoneNoneNoneNoneNoneNone NoneNoneNoneNoneNoneNoneNone NoneNoneNoneNoneNoneNoneNoneNoneNone

Figure 11 | To endow the capability of processing widely crawled PDFs (multilingual data), we train our model with OCR capabilities for nearly 100 languages. Minority language documents can also support both layout and non-layout outputs through different prompts.

4.3.3. General vision understanding

We also provide DeepSeek-OCR with a certain degree of general image understanding capabilities. The related visualization results are shown in Figure 12.

====================================================================== PAGE 18

Figure 1 Description: The image is a collage of four separate photographs, each depicting different scenes related to the theme of “image>nLocate 11-2= in the image.”

In the top left photograph, there is a classroom setting with a teacher and students. The teacher is standing in front of a whiteboard, which has a grid pattern and the text “11-2” written on it. The students are seated at desks, facing the teacher. The classroom has a green chalkboard on the left side and a whiteboard on the right side. The students are wearing school uniforms, and there are books and papers on the desks.

The top right photograph shows a group of children outdoors, with a green chalkboard in the background. The children are wearing school uniforms, and there are trees and a clear sky in the background.

The bottom left photograph is a close-up of a whiteboard with a grid pattern and the text “11-2” written on it. The whiteboard is mounted on a wall, and there is a window with a view of the outside.

The bottom right photograph is a close-up of a whiteboard with a grid pattern and the text “11-2” written on it. The whiteboard is mounted on a wall, and there is a window with a view of the outside.

The image also contains text in Chinese, which translates to “image>nLocate 11-2= in the image.” and “image>nLocate teacher in the image.”

The image is a visual representation of the concept of “image>nLocate 11-2= in the image,” which is a method used to locate specific elements within an image. The photographs illustrate this concept by showing different classroom settings and outdoor scenes, where the “image>nLocate 11-2=” method is applied.

Figure 12 | We retain DeepSeek-OCR’s capabilities in general visual understanding, mainly including image description, object detection, grounding, etc. Meanwhile, due to the inclusion of text-only data, DeepSeek-OCR’s language capabilities are also retained. Note that since we do not include SFT (Supervised Fine-Tuning) stage, the model is not a chatbot, and some capabilities need completion prompts to be activated.

5. Discussion

Our work represents an initial exploration into the boundaries of vision-text compression, investigating how many vision tokens are required to decode (N) text tokens. The preliminary results are encouraging: DeepSeek- OCR achieves near- lossless OCR compression at approximately (10 \times) ratios, while (20 \times) compression still retains (60\%) accuracy. These findings suggest promising directions for future applications, such as implementing optical processing for dialogue histories beyond (k) rounds in multi- turn conversations to achieve (10 \times) compression efficiency.

====================================================================== PAGE 19

Figure 1 Description:

• Memory:

  • Crystal Clear: Just happened (1 hour)
  • Very Clear: 1 day
  • Clear: 1 week
  • Blurry: 1 month
  • Very Blurry: Almost Gone (1 year)

• Vision:

  • Crystal Clear: 10cm
  • Very Clear: 50cm
  • Clear: 1m
  • Blurry: 3m
  • Very Blurry: Almost Gone (20m)

• Text:

  • Crystal Clear: Text token
  • Very Clear: Text token
  • Clear: Gundam
  • Blurry: Large
  • Very Blurry: Base
  • Almost Gone: Small
  • Tiny: Tiny

• Resolution:

  • Crystal Clear: Resolution
  • Very Clear: Resolution
  • Clear: Resolution
  • Blurry: Resolution
  • Very Blurry: Resolution
  • Almost Gone: Resolution
  • Tiny: Resolution

• Time:

  • Crystal Clear: Time →
  • Very Clear: Time →
  • Clear: Time →
  • Blurry: Time →
  • Very Blurry: Time →
  • Almost Gone: Time →

• Distance:

  • Crystal Clear: Distance ↑
  • Very Clear: Distance ↑
  • Clear: Distance ↑
  • Blurry: Distance ↑
  • Very Blurry: Distance ↑
  • Almost Gone: Distance ↑
  • Tiny: Distance ↑

The figure illustrates the progression of clarity and distance over time for different senses (memory, vision, and text) and resolutions. It shows how clarity and distance change from just happening to almost gone over time, and how these changes are reflected in the resolution of the text.

Figure 13 | Forgetting mechanisms constitute one of the most fundamental characteristics of human memory. The contexts optical compression approach can simulate this mechanism by rendering previous rounds of historical text onto images for initial compression, then progressively resizing older images to achieve multi-level compression, where token counts gradually decrease and text becomes increasingly blurred, thereby accomplishing textual forgetting.

For older contexts, we could progressively downsizing the rendered images to further reduce token consumption. This assumption draws inspiration from the natural parallel between human memory decay over time and visual perception degradation over spatial distance—both exhibit similar patterns of progressive information loss, as shown in Figure 13. By combining these mechanisms, contexts optical compression method enables a form of memory decay that mirrors biological forgetting curves, where recent information maintains high fidelity while distant memories naturally fade through increased compression ratios.

While our initial exploration shows potential for scalable ultra- long context processing, where recent contexts preserve high resolution and older contexts consume fewer resources, we acknowledge this is early- stage work that requires further investigation. The approach suggests a path toward theoretically unlimited context architectures that balance information retention with computational constraints, though the practical implications and limitations of such vision- text compression systems warrant deeper study in future research.

6. Conclusion

In this technical report, we propose DeepSeek- OCR and preliminarily validate the feasibility of contexts optical compression through this model, demonstrating that the model can effectively decode text tokens exceeding 10 times the quantity from a small number of vision tokens. We believe this finding will facilitate the development of VLMs and LLMs in the future. Additionally, DeepSeek- OCR is a highly practical model capable of large- scale pretraining data production, serving as an indispensable assistant for LLMs. Of course, OCR alone is insufficient to fully validate true context optical compression and we will conduct digital- optical text interleaved pretraining, needle- in- a- haystack testing, and other evaluations in the future. From another perspective, optical contexts compression still offers substantial room for research and improvement, representing a promising new direction.

====================================================================== PAGE 20

References

[1] Marker. URL https://github.com/datalab– to/marker.

[2] Mathpix. URL https://mathpix.com/.

[3] Ocrflux, 2025. URL https://github.com/chatdoc– com/OCRFlux.

[4] G. AI. Gemini 2.5- pro, 2025. URL https://gemini.google.com/.

[5] S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y. Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y. Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin. Qwen2.5- vl technical report. arXiv preprint arXiv:2502.13923, 2025.

[6] L. Blecher, G. Cucurull, T. Scialom, and R. Stojnic. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418, 2023.

[7] J. Chen, L. Kong, H. Wei, C. Liu, Z. Ge, L. Zhao, J. Sun, C. Han, and X. Zhang. Onechart: Purify the chart structural extraction via one auxiliary token. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 147- 155, 2024.

[8] Z. Chen, W. Wang, H. Tian, S. Ye, Z. Gao, E. Cui, W. Tong, K. Hu, J. Luo, Z. Ma, et al. How far are we to gpt- 4v? closing the gap to commercial multimodal models with open- source suites. arXiv preprint arXiv:2404.16821, 2024.

[9] C. Cui, T. Sun, M. Lin, T. Gao, Y. Zhang, J. Liu, X. Wang, Z. Zhang, C. Zhou, H. Liu, et al. Paddleocr 3.0 technical report. arXiv preprint arXiv:2507.05595, 2025.

[10] M. Dehghani, J. Djolonga, B. Mustafa, P. Padlewski, J. Heek, J. Gilmer, A. Steiner, M. Caron, R. Geirhos, I. Alabdulmohsin, et al. Patch n’ pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36:3632- 3656, 2023.

[11] H. Feng, S. Wei, X. Fei, W. Shi, Y. Han, L. Liao, J. Lu, B. Wu, Q. Liu, C. Lin, et al. Dolphin: Document image parsing via heterogeneous anchor prompting. arXiv preprint arXiv:2505.14059, 2025.

[12] Y. Goyal, T. Khot, D. Summers- Stay, D. Batra, and D. Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904- 6913, 2017.

[13] J. Gu, X. Meng, G. Lu, L. Hou, N. Minzhe, X. Liang, L. Yao, R. Huang, W. Zhang, X. Jiang, et al. Wukong: A 100 million large- scale chinese cross- modal pre- training benchmark. Advances in Neural Information Processing Systems, 35:26418- 26431, 2022.

[14] High- flyer. HAI- LLM: Efficient and lightweight training tool for large models, 2023. URL https://www.high– flyer.cn/en/blog/hai- llm.

[15] S. Iyer, X. V. Lin, R. Pasunuru, T. Mihaylov, D. Simig, P. Yu, K. Shuster, T. Wang, Q. Liu, P. S. Koura, et al. Opt- iml: Scaling language model instruction meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022.

[16] S. Kazemzadeh, V. Ordonez, M. Matten, and T. Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787- 798, 2014.

====================================================================== PAGE 21

[17] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.- Y. Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023.

[18] Z. Li, Y. Liu, Q. Liu, Z. Ma, Z. Zhang, S. Zhang, Z. Guo, J. Zhang, X. Wang, and X. Bai. Monkeyocr: Document parsing with a structure-recognition- relation triplet paradigm. arXiv preprint arXiv:2506.05218, 2025.

[19] A. Liu, B. Feng, B. Wang, B. Wang, B. Liu, C. Zhao, C. Dengr, C. Ruan, D. Dai, D. Guo, et al. Deepseek- v2: A strong, economical, and efficient mixture- of- experts language model. arXiv preprint arXiv:2405.04434, 2024.

[20] A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, et al. Deepseek- v3 technical report. arXiv preprint arXiv:2412.19437, 2024.

[21] C. Liu, H. Wei, J. Chen, L. Kong, Z. Ge, Z. Zhu, L. Zhao, J. Sun, C. Han, and X. Zhang. Focus anywhere for fine- grained multi- page document understanding. arXiv preprint arXiv:2405.14295, 2024.

[22] I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.

[23] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In ICLR, 2019.

[24] A. Masry, D. X. Long, J. Q. Tan, S. Joty, and E. Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022.

[25] A. Nassar, A. Marafioti, M. Omenetti, M. Lysak, N. Livathinos, C. Auer, L. Morin, R. T. de Lima, Y. Kim, A. S. Gurbuz, et al. Smoldocling: An ultra- compact vision- language model for end- to- end multi- modal document conversion. arXiv preprint arXiv:2503.11576, 2025.

[26] OpenAI. Gpt- 4 technical report, 2023.

[27] L. Ouyang, Y. Qu, H. Zhou, J. Zhu, R. Zhang, Q. Lin, B. Wang, Z. Zhao, M. Jiang, X. Zhao, et al. Omnidocbench: Benchmarking diverse pdf document parsing with comprehensive annotations. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24838- 24848, 2025.

[28] J. Poznanski, A. Rangapur, J. Borchardt, J. Dunkelberger, R. Huff, D. Lin, C. Wilhelm, K. Lo, and L. Soldaini. olmocr: Unlocking trillions of tokens in pdfs with vision language models. arXiv preprint arXiv:2502.18443, 2025.

[29] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748- 8763. PMLR, 2021.

[30] Rednote. dots.ocr, 2025. URL https://github.com/rednote– hilab/dots.ocr.

[31] C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmarczyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki. Laion- 400m: Open dataset of clip- filtered 400 million image- text pairs. arXiv preprint arXiv:2111.02114, 2021.

====================================================================== PAGE 22

[32] A. Singh, V. Natarajan, M. Shah, Y. Jiang, X. Chen, D. Batra, D. Parikh, and M. Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317- 8326, 2019.

[33] T. Sun, C. Cui, Y. Du, and Y. Liu. Pp- dcolayout: A unified document layout detection model to accelerate large- scale data construction. arXiv preprint arXiv:2503.17213, 2025.

[34] B. Wang, C. Xu, X. Zhao, L. Ouyang, F. Wu, Z. Zhao, R. Xu, K. Liu, Y. Qu, F. Shang, et al. Mineru: An open- source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024.

[35] P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, et al. Qwen2- vl: Enhancing vision- language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024.

[36] H. Wei, L. Kong, J. Chen, L. Zhao, Z. Ge, J. Yang, J. Sun, C. Han, and X. Zhang. Vary: Scaling up the vision vocabulary for large vision- language model. In European Conference on Computer Vision, pages 408- 424. Springer, 2024.

[37] H. Wei, L. Kong, J. Chen, L. Zhao, Z. Ge, E. Yu, J. Sun, C. Han, and X. Zhang. Small language model meets with reinforced vision vocabulary. arXiv preprint arXiv:2401.12503, 2024.

[38] H. Wei, C. Liu, J. Chen, J. Wang, L. Kong, Y. Xu, Z. Ge, L. Zhao, J. Sun, Y. Peng, et al. General ocr theory: Towards ocr- 2.0 via a unified end- to- end model. arXiv preprint arXiv:2409.01704, 2024.

[39] H. Wei, Y. Yin, Y. Li, J. Wang, L. Zhao, J. Sun, Z. Ge, X. Zhang, and D. Jiang. Slow perception: Let’s perceive geometric figures step- by- step. arXiv preprint arXiv:2412.20631, 2024.

[40] Z. Wu, X. Chen, Z. Pan, X. Liu, W. Liu, D. Dai, H. Gao, Y. Ma, C. Wu, B. Wang, et al. Deepseek- vl2: Mixture- of- experts vision- language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024.

[41] W. Yu, Z. Yang, L. Li, J. Wang, K. Lin, Z. Liu, X. Wang, and L. Wang. Mm- vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023.

[42] J. Zhu, W. Wang, Z. Chen, Z. Liu, S. Ye, L. Gu, H. Tian, Y. Duan, W. Su, J. Shao, et al. Internvl3: Exploring advanced training and test- time recipes for open- source multimodal models. arXiv preprint arXiv:2504.10479, 2025.