← 首页 | 导读 | 详细解读

DeepSeek-OCR: Contexts Optical Compression

DeepSeek-OCR：上下文光学压缩

📄 arXiv: 2510.18234📅 2025-10-22英文 PDF中文 PDF

翻译进度 45 / 45 段 (100%)

中文摘要

DeepSeek-OCR 采用光学上下文压缩技术，通过 2D 光学映射实现高效的文档压缩和 OCR 识别。该模型能够在保持语义完整性的前提下大幅压缩文档图像信息，实现超高速的文档分析和文本提取。在扫描文档、合同、论文等场景下表现出优异的识别准确率和速度。

阅读模式

左侧英文原版 · 右侧中文 PDF（A4 双栏排版）· 可分别滚动对照

English Original

中文翻译 PDF

中文翻译

📑 目录

DeepSeek-OCR: Contexts Optical Compression DeepSeek-OCR: Contexts Optical Compression DeepSeek-OCR: Contexts Optical Compression DeepSeek-OCR: Contexts Optical Compression DeepSeek-OCR: Contexts Optical Compression DeepSeek-OCR: Contexts Optical Compression DeepSeek-OCR: Contexts Optical Compression 1 Introduction 1 Introduction 1 Introduction 2 Related Works 2 Related Works 2.1 Typical Vision Encoders in VLMs 2.2 End-to-end OCR Models 3 Methodology 3 Methodology 3 Methodology 3 Methodology 3 Methodology 3 Methodology 3 Methodology 3 Methodology 3 Methodology 3 Methodology 3.1 Architecture 3.2 DeepEncoder 3.2 DeepEncoder 3.2 DeepEncoder 3.3 The MoE Decoder 3.4 Data Engine 3.4 Data Engine 3.4 Data Engine 3.5 Training Pipelines 3.5 Training Pipelines 3.5 Training Pipelines 4 Evaluation 4 Evaluation 4 Evaluation 4 Evaluation 4.1 Vision-text Compression Study 4.2 OCR Practical Performance 4.3 Qualitative Study 4.3 Qualitative Study 5 Discussion 6 Conclusion

DeepSeek-OCR: Contexts Optical Compression

Haoran Wei, Yaofeng Sun, Yukun Li, DeepSeek-AI 摘要：我们提出DeepSeek-OCR作为通过光学2D映射压缩长上下文的可行性初步探索。DeepSeek-OCR由两个组件组成：DeepEncoder和作为解码器的DeepSeek3B-MoE-A570M。具体来说，DeepEncoder作为核心引擎，旨在高分辨率输入下保持低激活状态，同时实现高效的视觉token压缩。我们的方法证明了通过视觉token进行光学压缩的可行性，为未来多模态模型设计提供了新的思路。

[原文]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 compr...

DeepSeek-OCR: Contexts Optical Compression

文本信息。一张包含文档文本的图像可以用远少于等价数字文本的 token 表示丰富的信息，这表明通过 vision token 进行的光学压缩可能实现更高的压缩比。这一洞见促使我们从以 LLM 为中心的视角重新审视视觉-语言模型（VLM），关注 vision encoder 如何提升 LLM 处理文本信息的效率，而非人类擅长的基础 VQA [12, 24, 32, 41, 16]。OCR 任务作为连接视觉与语言的中间模态……

[原文]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 , 24 , 32 , 41 , 16 ] what humans excel at. OCR tasks, as an intermediate modality bridging vision and language, ...

DeepSeek-OCR: Contexts Optical Compression

window attention 组件处理大量 vision token，而 compressor 在它们进入 dense global attention 组件之前减少 vision token 数量，从而实现有效的内存与 token 压缩。第三，我们基于 DeepEncoder 和 DeepSeek3B-MoE [19, 20] 开发 DeepSeek-OCR。如图 LABEL:fig:omni 所示，它在 OmniDocBench 上以端到端模型中最少的 vision token 实现了 state-of-the-art 性能。此外，我们为模型配备了解析图表、化学公式、简单几何图形和自然图像的能力……

[原文]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 LABEL:fig:omni , 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 t...

DeepSeek-OCR: Contexts Optical Compression

第一种是由 Vary [36] 代表的双塔架构，利用并行的 SAM [17] encoder 增加 visual vocabulary 参数以处理高分辨率图像。虽然该方案提供可控的参数与 activation memory，但存在显著缺陷：需要双重图像预处理，使部署复杂化，并在训练时使 encoder pipeline parallelism 面临挑战。第二种是以 InternVL2.0 [8] 为代表的 tile-based 方法，通过将图像分割为小 tile 进行并行计算来降低……

[原文]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 ...

DeepSeek-OCR: Contexts Optical Compression

任务。GOT-OCR2.0 [38] 将 OCR2.0 的范围扩展至更多合成图像解析任务，并设计了兼顾性能与效率的 OCR 模型，进一步凸显了端到端 OCR 研究的潜力。此外，Qwen-VL 系列 [35]、InternVL 系列 [8] 及其众多衍生模型持续增强文档 OCR 能力，以探索 dense visual perception 的边界。然而，当前模型尚未解决的一个关键研究问题是：对于包含 1000 个单词的文档，至少需要多少 vision token……

[原文]sks. 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 ...

DeepSeek-OCR: Contexts Optical Compression

高分辨率下 activation 较低；3. vision token 数量少；4. 支持多种分辨率输入；5. 参数量适中。然而，如第 2.1 节所述，当前开源 encoder 无法完全满足所有这些条件。因此，我们自行设计了一种新型 vision encoder，命名为 DeepEncoder。图 4：为测试不同压缩比（需要不同数量的 vision token）下的模型性能，并增强 DeepSeek-OCR 的实用性，我们为其配置了多种分辨率模式。3.2.1 DeepEncoder 架构 DeepEncoder 主要由两个组件构成……

[原文]tivation 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. 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. 3.2.1 Architecture of DeepEncoder DeepEncoder mainly consists of two compone...

DeepSeek-OCR: Contexts Optical Compression

Note: The original text has `Pipline Models` and `End-to-end Models` as section headers within the table. I'll keep them as row headers or just translate them inline. I will format it cleanly. Also, `without layout: "\nFree OCR." to control the model’s output 尽管如此，输出格式仍无法完全匹配Fox基准测试，因此实际性能可能会略高于测试结果。如表2所示，在10倍压缩率以内，模型的解码精度可达约97%，这是一个非常令人鼓舞的结果。未来，通过文本到图像的方法，有望实现接近10倍的无损上下文压缩。当压缩率超过10倍时，性能开始下降，这可能有两方面原因：一是长文档的排版变得更加复杂；二是在512×512或640×640分辨率下，长文本会出现模糊现象。第一个问题可以通过将文本渲染至单一排版页面来解决，而我们认为第二个问题将成为遗忘机制的一个特征。当Token压缩率接近20倍时，我们发现精度仍可接近60%。这些结果表明，光学上下文压缩是一个极具前景且值得深入研究的方向。该方法不会带来额外开销，因为它可以复用VLM（视觉语言模型）的基础设施，而多模态系统本身就需要额外的视觉编码器。

[原文]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 "†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 table order Pipline Mode...

**4.2. OCR实际性能** DeepSeek-OCR不仅是一个实验性模型，它还具备强大的实际应用能力，可用于构建大语言模型（LLM）/视觉语言模型（VLM）的预训练数据。为量化OCR性能，我们在OmniDocBench [27]上对DeepSeek-OCR进行了测试，结果如表3所示。仅需100个视觉Token（640×640分辨率），DeepSeek-OCR的性能便超越了使用256个Token的GOT-OCR2.0 [38]；在400个Token（285个有效Token，1280×1280分辨率）的设定下，其表现与该基准测试上的当前最优方法持平。在使用少于800个Token（Gundam模式）的情况下，DeepSeek-OCR的性能优于需要近7000个视觉Token的MinerU2.0 [34]。这些结果表明，我们的DeepSeek 对于书籍和报告类文档，DeepSeek-OCR 仅需 100 个视觉 Token 即可实现良好的性能。结合第 4.1 节的分析，这可能是因为此类文档中的文本 Token 数量大多在 1,000 以内，意味着视觉 Token 的压缩比不超过 10 倍。对于报纸类文档，则需要使用 Gundam 甚至 Gundam-master 模式才能获得可接受的编辑距离，因为报纸中的文本 Token 数量高达 4,000 至 5,000，远超其他模式 10 倍的压缩能力。这些实验结果进一步揭示了上下文光学压缩的边界，有望为视觉语言模型（VLM）中的视觉 Token 优化、上下文压缩，以及大语言模型（LLM）中的遗忘机制研究提供有效参考。 DeepSeek-OCR 兼具版面分析与 OCR 2.0 能力，能够通过二次模型调用对文档内的图像进行进一步解析，我们将此特性称为“深度解析”。如图 7、8、9、10 所示，我们的模型仅需一个统一的提示词，即可对图表、几何图形、化学式甚至自然图像进行深度解析。12

1 Introduction

当前大语言模型（LLM）在处理长文本内容时面临显著的计算挑战，因为计算量随序列长度呈二次方增长。我们探索一种潜在解决方案：利用视觉模态作为文本信息的高效压缩媒介。一张包含文档文本的图像可以用远少于等价数字文本的 token 表示丰富的信息，表明通过 vision token 进行的光学压缩可能实现更高的压缩比。这一洞见促使我们从以 LLM 为中心的视角重新审视视觉-语言模型（VLM）……

[原文]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 a...

1 Introduction

一种即使在高分辨率输入下仍保持低 activation memory 和最少 vision token 的架构。它通过 16× 卷积 compressor 将 window attention 与 global attention encoder 组件串行连接。该设计确保 window attention 组件处理大量 vision token，而 compressor 在它们进入 dense global attention 组件之前减少 vision token 数量，实现有效的内存与 token 压缩。第三，我们基于 DeepEncoder 和 DeepSeek3B-MoE [19, 20] 开发 DeepSeek-OCR。如图 LABEL:fig:om……

[原文]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 LABEL:fig:om...

1 Introduction

当前开源VLM中常用的编码器，各自都存在各自的缺陷。这些缺陷包括计算复杂度高、内存占用大、token数量多等问题。我们的DeepEncoder旨在解决这些限制。这些限制包括计算复杂度高、内存占用大、token数量多等问题，限制了模型在实际部署中的效率。我们的DeepEncoder旨在解决这些限制，通过创新的架构设计实现高效的光学压缩。

[原文]ncoders commonly used in current open-source VLMs, all of which suffer from their respective deficiencies.

2 Related Works

2.1 VLM 中的典型 Vision Encoder 当前开源 VLM 采用三种主要类型的 vision encoder，如图 2 所示。第一种是由 Vary [36] 代表的双塔架构，利用并行的 SAM [17] encoder 增加 visual vocabulary 参数以处理高分辨率图像。虽然该方案提供可控的参数与 activation memory，但存在显著缺陷：需要双重图像预处理，使部署复杂化，并在训练时使 encoder pipeline parallelism 面临挑战。第二种是以 tile-ba……

2 Related Works

首先在 arXiv 上采用端到端框架进行学术论文 OCR，展示了模型处理 dense perception 任务的潜力。GOT-OCR2.0 [38] 将 OCR2.0 的范围扩展至更多合成图像解析任务，并设计了兼顾性能与效率的 OCR 模型，进一步凸显了端到端 OCR 研究的潜力。此外，Qwen-VL 系列 [35]、InternVL 系列 [8] 及其众多衍生模型持续增强文档 OCR 能力，以探索 dense visual perception 的边界。然而，一个关键……

[原文]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 crucia...

2.1 Typical Vision Encoders in VLMs

当前开源 VLM 采用三种主要类型的 vision encoder，如图 2 所示。第一种是由 Vary [36] 代表的双塔架构，利用并行的 SAM [17] encoder 增加 visual vocabulary 参数以处理高分辨率图像。虽然该方案提供可控的参数与 activation memory，但存在显著缺陷：需要双重图像预处理，使部署复杂化，并在训练时使 encoder pipeline parallelism 面临挑战。第二种是以 InternVL2.0 [8] 为代表的 tile-based 方法……

[原文]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....

2.2 End-to-end OCR Models

OCR，尤其是文档解析任务，一直是 image-to-text 领域的热点话题。随着 VLM 的发展，大量端到端 OCR 模型涌现，通过简化 OCR 系统从根本上改变了传统 pipeline 架构（需要独立的检测与识别专家模型）。Nougat [6] 首先在 arXiv 上采用端到端框架进行学术论文 OCR，展示了模型处理 dense perception 任务的潜力。GOT-OCR2.0 [38] 将 OCR2.0 的范围扩展至更多合成图像解析任务并设计……

[原文]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 desig...

3 Methodology

3.1 架构如图 3 所示，DeepSeek-OCR 采用统一的端到端 VLM 架构，由 encoder 和 decoder 组成。encoder（即 DeepEncoder）负责提取图像特征、tokenize 以及压缩 visual representation。decoder 基于 image token 和 prompt 生成所需结果。DeepEncoder 参数量约 380M，主要由 80M 的 SAM-base [17] 和 300M 的 CLIP-large [29] 串行连接构成。decoder 采用 3B MoE [19, 20] 架构，激活参数量约 570M。在……

3 Methodology

36] 并使用 2 层卷积模块对 vision token 进行 16× 下采样。每层卷积的 kernel size 为 3、stride 为 2、padding 为 1，通道数从 256 增至 1024。假设输入 1024×1024 图像，DeepEncoder 将其分割为 1024/16×1024/16=4096 个 patch token。由于 encoder 前半部分以 window attention 为主且仅 80M，activation 可接受。在进入 global attention 之前，4096 个 token 经过压缩模块，token 数量变为 4096/16=256，从而使整体 a……

[原文]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, the 4096 tokens go through the compression module and the token count becomes 4096/16=256, thus making the overall a...

3 Methodology

）分别。由于 Tiny 和 Small 模式分辨率相对较小，为避免浪费 vision token，图像通过直接 resize 原始形状处理。对于 Base 和 Large 模式，为保留原始图像宽高比，图像被 padding 至对应尺寸。padding 后，有效 vision token 数量小于实际 vision token 数量，计算公式为：N_valid = ⌈N_actual × [1 − ((max(w,h) − min(w,h)) / max(w,h))]⌉，N_{valid}=⌈N_{actual}……

[原文]) 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 v a l i d = ⌈ N a c t u a l × [ 1 − ( ( m a x ( w , h ) − m i n ( w , h ) ) / ( m a x ( w , h ) ) ) ] ⌉ N_{valid}=\lceil N_{actual}...

3 Methodology

Gundam-master 的分辨率过大，与其一起训练会拖慢整体训练速度。3.3 MoE Decoder 我们的 decoder 使用 DeepSeekMoE [19, 20]，具体为 DeepSeek-3B-MoE。推理时，模型激活 64 个 routed expert 中的 6 个和 2 个 shared expert，激活参数量约 570M。3B DeepSeekMoE 非常适合以领域为中心（对我们而言是 OCR）的 VLM 研究，因为它获得 3B 模型的表达能力，同时享有 500M 小模型的推理效率。decoder 从 com…… 的压缩 latent vision token 重建原始文本表示。

[原文]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 com...

3 Methodology

标注展示。我们将 ground truth 格式化为交错的 layout 与 text 格式，每段文本前标注其在原图中的坐标和标签。所有坐标归一化至 1000 个 bin。3.4.1 OCR 1.0 数据文档数据是 DeepSeek-OCR 的首要任务。我们从互联网收集 3000 万页涵盖约 100 种语言的多样化 PDF 数据，其中中英文约 2500 万页，其他语言约 500 万页。对此数据，我们创建两类 ground truth：粗标注和细标注。粗标注……

[原文]tions 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. 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...

3 Methodology

文和英文。与文档 OCR 类似，自然场景 OCR 也可通过 prompt 控制是否输出检测框。3.4.2 OCR 2.0 数据遵循 GOT-OCR2.0 [38]，我们将图表、化学公式和平面几何解析数据称为 OCR 2.0 数据。对于图表数据，遵循 OneChart [7]，我们使用 pyecharts 和 matplotlib 渲染 1000 万张图像，主要包括常用的折线图、柱状图、饼图和复合图表。我们将图表解析定义为 image-to-HTML-table 转换任务，如图 LABEL:fig:demo2-1 所示。对于化学公式，我们使用 PubChem 的 SMILES 格式作为数据……

[原文]nese 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 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 LABEL:fig:demo2-1 . For chemical formulas, we utilize SMILES format from PubChem as the data ...

3 Methodology

并非通用 VLM 模型，这部分数据仅占总数据的 20%。引入此类数据主要是保留通用视觉接口，以便对我们模型和通用视觉任务感兴趣的研究者在未来方便推进工作。3.4.4 纯文本数据为确保模型的语言能力，我们引入 10% 的内部纯文本预训练数据，所有数据处理至 8192 token 长度，这也是 DeepSeek-OCR 的序列长度。综上，训练 DeepSeek-OCR 时，OCR 数据占 70%，通用视觉……

[原文]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 ...

3 Methodology

，同时将 CLIP 部分作为 input embedding layer 置于 PP1 并解冻权重进行训练。对于语言模型部分，由于 DeepSeek3B-MoE 有 12 层，我们在 PP2 和 PP3 各放置 6 层。我们使用 20 个节点（每个 8 张 A100-40G GPU）训练，data parallelism（DP）为 40，global batch size 为 640。我们使用 AdamW 优化器和 step-based scheduler，初始学习率为 3e-5。纯文本数据的训练速度为 90B tokens/day，多模态数据为 70B tokens/day。表 2：我们测试 DeepSeek-OCR 的 vision-tex……

[原文], 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-tex...

3 Methodology

all text formula table order Pipeline Models Dolphin [11] - 0.356 0.352 0.465 0.258 0.35 0.44 0.44 0.604 0.367 0.351 Marker [1] - 0.296 0.085 0.374 0.609 0.116 0.497 0.293 0.688 0.678 0.329 Mathpix [2] - 0.191 0.105 0.306 0.243 0.108 0.364 0.381 0.454 0.32 0.30 MinerU-2.1.1 [34] - 0.162 0.072 0.313 0.166 0.097 0.244 0.111 0.581 0.15 0.136 MonkeyOCR-1.2B [18] - 0.154 0.062 0.295 0.164 0.094 0.263 0.179 0.464 0.168 0.243 PPstructure-v3 [9] - 0.152 0.073 0.295 0.162 0.077 0.223 0.136 0.535 0.111 0.11 End-to-end Models Nougat [6] 2352 0.452 0.365 0.488 0.572 0.382 0.973 0.998 0.941 1……

[原文]all text formula table order Pipline Models Dolphin [ 11 ] - 0.356 0.352 0.465 0.258 0.35 0.44 0.44 0.604 0.367 0.351 Marker [ 1 ] - 0.296 0.085 0.374 0.609 0.116 0.497 0.293 0.688 0.678 0.329 Mathpix [ 2 ] - 0.191 0.105 0.306 0.243 0.108 0.364 0.381 0.454 0.32 0.30 MinerU-2.1.1 [ 34 ] - 0.162 0.072 0.313 0.166 0.097 0.244 0.111 0.581 0.15 0.136 MonkeyOCR-1.2B [ 18 ] - 0.154 0.062 0.295 0.164 0.094 0.263 0.179 0.464 0.168 0.243 PPstructure-v3 [ 9 ] - 0.152 0.073 0.295 0.162 0.077 0.223 0.136 0.535 0.111 0.11 End-to-end Models Nougat [ 6 ] 2352 0.452 0.365 0.488 0.572 0.382 0.973 0.998 0.941 1....

3 Methodology

这不是一般的VLM模型，这部分数据仅占总数据的20%。我们引入此类数据主要是为了保留通用视觉接口，使对我们的模型和通用视觉任务感兴趣的研究人员能够方便地推进他们的工作。3.4.4 纯文本数据：为确保模型的语言能力，我们引入了10%的纯文本预训练数据。

[原文]0.269 0.134 0.062 0.181 0.097 0.432 0.089 0.103 Gundam-M †200dpi 1853 0.123 0.049 0.242 0.147 0.056 0.157 0.087 0.377 0.08 0.085

3.1 Architecture

如图 3 所示，DeepSeek-OCR 采用统一的端到端 VLM 架构，由 encoder 和 decoder 组成。encoder（即 DeepEncoder）负责提取图像特征、tokenize 以及压缩 visual representation。decoder 基于 image token 和 prompt 生成所需结果。DeepEncoder 参数量约 380M，主要由 80M 的 SAM-base [17] 和 300M 的 CLIP-large [29] 串行连接构成。decoder 采用 3B MoE [19, 20] 架构，激活参数量约 570M。在以下段落中……

[原文]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 paragr...

3.2 DeepEncoder

为探索 contexts optical compression 的可行性，我们需要具备以下特征的 vision encoder：1. 能处理高分辨率；2. 高分辨率下 activation 较低；3. vision token 数量少；4. 支持多种分辨率输入；5. 参数量适中。然而，如第 2.1 节所述，当前开源 encoder 无法完全满足所有这些条件。因此，我们自行设计了一种新型 vision encoder，命名为 DeepEncoder。图 4：为测试不同压缩比（需要不同数量的 vision token）下的模型性能并 en……

[原文]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. Figure 4 : To test model performance under different compression ratios (requiring different numbers of vision tokens) and en...

3.2 DeepEncoder

假设我们有一张包含1000个光学字符的图像，并希望测试解码需要多少个视觉Token。这要求模型支持可变数量的视觉Token。也就是说，DeepEncoder需要支持多种分辨率。我们通过位置编码的动态插值来满足上述要求，并设计了多种分辨率模式进行模型的联合训练，从而使单个 DeepSeek-OCR 模型具备支持多分辨率的能力。如图4所示，DeepEncoder 主要支持两种输入模式：原生分辨率（native resolution）和动态分辨率（dynamic resolution）。每种模式均包含多个子模式。原生分辨率支持四个子模式：Tiny、Small、Base 和 Large，对应的分辨率和 token 数量分别为 512×512（64）、640×640（100）、1024×1024（256）和 1280×1280（400）。由于 Tiny 和 Small 模式的分辨率相对较小，为避免浪费视觉 token，图像通过直接缩放原始尺寸进行处理。对于 Base 和 Large 模式，为保持原始图像的宽高比，图像会被填充至对应尺寸。填充后，有效视觉 token 的数量将少于实际视觉 token 的数量，计算公式为：𝑁𝑣𝑎𝑙𝑖𝑑= ⌈𝑁𝑎𝑐𝑡𝑢𝑎𝑙× [1 −((𝑚𝑎𝑥(𝑤, ℎ) −𝑚𝑖𝑛(𝑤, ℎ))/(𝑚𝑎𝑥(𝑤, ℎ)))]⌉ (1)，其中 𝑤 和 ℎ 分别表示原始输入图像的宽度和高度。6

[原文]n=6 R=1-(H-W)/W W:512||640 H:512||640 Mode: Tiny||Small W:1024||1280 Resize Padding H:1024||1280 Mode: Base||Large Token: 64||100 Token: 256||400 Valid: (256||400)×R + 640||1024 640||1024 W:1024||1280 H:1024||1280 Mode: Gundam||Gundam (Master) Token: n×(100||256) + (256||400) Valid: n×(100||256) + (256||400)×R n∈[2:9] 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 cou...

3.2 DeepEncoder

输入图像。动态分辨率可由两种 native resolution 组合构成。例如，Gundam 模式由 n×640×640 个 tile（local view）和 1024×1024 的 global view 组成。分块方法遵循 InternVL2.0 [8]。支持动态分辨率主要是出于应用考虑，尤其是超高分辨率输入（如报纸图像）。Tiling 是一种 secondary window attention 形式，可进一步有效降低 activation memory。值得注意的是，由于我们的 native resolution 相对较大，图像不会被过度碎片化 un……

[原文]nput image. 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 un...

3.3 The MoE Decoder

我们的 decoder 使用 DeepSeekMoE [19, 20]，具体为 DeepSeek-3B-MoE。推理时，模型激活 64 个 routed expert 中的 6 个和 2 个 shared expert，激活参数量约 570M。3B DeepSeekMoE 非常适合以领域为中心（对我们而言是 OCR）的 VLM 研究，因为它获得 3B 模型的表达能力，同时享有 500M 小模型的推理效率。decoder 从 DeepEncoder 压缩的 latent vision token 重建原始文本表示，形式为：f_dec : ℝ^{n×d_latent} → ℝ^{N×d_text}；𝐗̂ = f_dec(𝐙)，其中 n ≤ N，f_{tex……

[原文]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 dec : ℝ n × d latent → ℝ N × d text ; 𝐗 ^ = f dec ( 𝐙 ) where n ≤ N f_{\tex...

3.4 Data Engine

我们为 DeepSeek-OCR 构建了复杂多样的训练数据，包括 OCR 1.0 数据（主要包括场景图像 OCR 和文档 OCR 等传统 OCR 任务）、OCR 2.0 数据（主要包括图表、化学公式和平面几何解析等复杂人工图像的解析任务）、以及通用视觉数据（主要用于向 DeepSeek-OCR 注入一定的通用图像理解能力并保留通用视觉接口）。(a) OCR 1.0 细标注展示。我们将 ground truth 格式化为交错的 layout 与 text fo……

[原文]We constructe 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. (a) OCR 1.0 fine annotations display. We format the ground truth into an interleaved layout and text fo...

3.4 Data Engine

图像-文本对示例见图 6(a)。我们还收集 300 万 Word 数据，通过直接提取内容构建无 layout 的高质量图像-文本对。该数据主要对公式和 HTML 格式表格有益。此外，我们选取部分开源数据 [28, 37] 作为补充。对于自然场景 OCR，模型主要支持中英文。图像数据来源为 LAION [31] 和 Wukong [13]，使用 PaddleOCR [9] 标注，中英文各 1000 万样本。与文档 OCR 类似，自然场景 OCR 也可通过 prompt 控制是否……

[原文]ation image-text pairs can be seen in Figure 6(a) . 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 ...

3.4 Data Engine

如线段、端点坐标、线段类型等，以提升可读性。每条线段采用 Slow Perception [39] 方式编码。3.4.3 通用视觉数据 DeepEncoder 可受益于 CLIP 的预训练增益，且有足够参数融入通用视觉知识。因此，我们也为 DeepSeek-OCR 准备了相应数据。遵循 DeepSeek-VL2 [40]，我们为 caption、detection 和 grounding 等任务生成相关数据。需注意 DeepSeek-OCR 并非通用 VLM 模型，这部分数据仅占总数据的 20%……

[原文]such as line segments, endpoint coordinates, line segment types, etc., for better readability. Each line segment is encoded using the Slow Perception [ 39 ] manner. 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...

3.5 Training Pipelines

我们的训练 pipeline 非常简洁，主要包括两个阶段：a) 独立训练 DeepEncoder；b) 训练 DeepSeek-OCR。注意 Gundam-master 模式是在预训练 DeepSeek-OCR 模型上继续训练 600 万采样数据获得的。由于训练协议与其他模式相同，此处省略详细描述。3.5.1 训练 DeepEncoder 遵循 Vary [36]，我们使用 compact language model [15] 和 next token prediction 框架训练 DeepEncoder。此阶段使用上述全部 OCR 1.0 和 2.0 数据，以及 a……

[原文]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. 3.5.1 Training DeepEncoder Following 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 a...

3.5 Training Pipelines

21] 基准。Text tokens 表示使用 DeepSeek-OCR tokenizer 对 ground truth 文本 tokenize 后的 token 数量。Vision Tokens=64 或 100 分别表示 DeepEncoder 将输入图像 resize 至 512×512 和 640×640 后输出的 vision token 数量。Text Tokens Vision Tokens=64 Vision Tokens=100 Precision Compression Precision Compression Pages 600-700 96.5% 10.5× 98.5% 6.7× 7 700-800 93.8% 11.8× 97.3% 7.5× 28 800-900 83.8% 13.2× 96.8% 8.5× 28 900-1000 85.9% 15.1× 96.8% 9.7× 14 1000……

[原文]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 × \times 512 and 640 × \times 640. Text Tokens Vision Tokens =64 Vision Tokens=100 Precision Compression Precision Compression Pages 600-700 96.5% 10.5 × \times 98.5% 6.7 × \times 7 700-800 93.8% 11.8 × \times 97.3% 7.5 × \times 28 800-900 83.8% 13.2 × \times 96.8% 8.5 × \times 28 900-1000 85.9% 15.1 × \times 96.8% 9.7 × \times 14 1000...

3.5 Training Pipelines

07 0.522 InternVL2-76B [8] 6790 0.44 0.353 0.543 0.547 0.317 0.443 0.29 0.701 0.555 0.228 Qwen2.5-VL-7B [5] 3949 0.316 0.151 0.376 0.598 0.138 0.399 0.243 0.5 0.627 0.226 OLMOCR [28] 3949 0.326 0.097 0.455 0.608 0.145 0.469 0.293 0.655 0.652 0.277 GOT-OCR2.0 [38] 256 0.287 0.189 0.360 0.459 0.141 0.411 0.315 0.528 0.52 0.28 OCRFlux-3B [3] 3949 0.238 0.112 0.447 0.269 0.126 0.349 0.256 0.716 0.162 0.263 GPT4o [26] - 0.233 0.144 0.425 0.234 0.128 0.399 0.409 0.606 0.329 0.251 InternVL3-78B [42] 6790 0.218 0.117 0.38 0.279 0.095 0.296 0.21 0.533 0.282 0.161 Qwen2.5-VL-72B [5] 3949……

[原文]07 0.522 InternVL2-76B [ 8 ] 6790 0.44 0.353 0.543 0.547 0.317 0.443 0.29 0.701 0.555 0.228 Qwen2.5-VL-7B [ 5 ] 3949 0.316 0.151 0.376 0.598 0.138 0.399 0.243 0.5 0.627 0.226 OLMOCR [ 28 ] 3949 0.326 0.097 0.455 0.608 0.145 0.469 0.293 0.655 0.652 0.277 GOT-OCR2.0 [ 38 ] 256 0.287 0.189 0.360 0.459 0.141 0.411 0.315 0.528 0.52 0.28 OCRFlux-3B [ 3 ] 3949 0.238 0.112 0.447 0.269 0.126 0.349 0.256 0.716 0.162 0.263 GPT4o [ 26 ] - 0.233 0.144 0.425 0.234 0.128 0.399 0.409 0.606 0.329 0.251 InternVL3-78B [ 42 ] 6790 0.218 0.117 0.38 0.279 0.095 0.296 0.21 0.533 0.282 0.161 Qwen2.5-VL-72B [ 5 ] 3949...

4 Evaluation

4.1 视觉-文本压缩研究我们选取 Fox [21] 基准验证 DeepSeek-OCR 对富文本文档的压缩-解压缩能力，以初步探索 contexts optical compression 的可行性与边界。我们使用 Fox 的英文文档部分，用 DeepSeek-OCR tokenizer（词表规模约 129k）对 ground truth 文本进行 tokenize，选取 600-1300 token 的文档进行测试，恰好为 100 页。由于 text token 数量不大，我们只需在 Tiny 和 Small 模式测试性能，其中 Tiny mo……

4 Evaluation

encoder。表 4：OmniDocBench 上不同文档类别的编辑距离。结果表明，某些文档类型仅需 64 或 100 个 vision token 即可取得良好性能，而其他类型则需要 Gundam 模式。Mode Type 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 0.32 Small 0.085 0.111 0.079 0.147 0.171 0.107 0.131 0.187 0.744 0.205 Base 0.037 0.08 0.027 0.1 0.13 0.073 0.052 0.176 0.645 0.156 Large 0.038 0.108 0.022 0.084 0.109 0.06 0.053 0.155 0.353 0.117 Gunda……

[原文]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. Mode Type 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 0.32 Small 0.085 0.111 0.079 0.147 0.171 0.107 0.131 0.187 0.744 0.205 Base 0.037 0.08 0.027 0.1 0.13 0.073 0.052 0.176 0.645 0.156 Large 0.038 0.108 0.022 0.084 0.109 0.06 0.053 0.155 0.353 0.117 Gunda...

4 Evaluation

不超过 10×。对于报纸，需要 Gundam 甚至 Gundam-master 模式才能达到可接受的编辑距离，因为报纸的 text token 为 4-5,000，远超其他模式的 10× 压缩。这些实验结果进一步展示了 contexts optical compression 的边界，可为 VLM 中 vision token 优化、context compression 以及 LLM 遗忘机制等研究提供有效参考。4.3 定性研究 4.3.1 深度解析 DeepSeek-OCR 同时具备 layout 和 OCR 2.0 能力，使其能够……

[原文]o does not exceed 10 × \times . 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 10 × \times 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...

4 Evaluation

互联网上的 PDF 数据不仅包含中英文，还有大量多语言数据，这在训练 LLM 时也至关重要。对于 PDF 文档，DeepSeek-OCR 可处理近 100 种语言。与中英文文档类似，多语言数据也支持 layout 和非 layout 两种 OCR 格式。可视化结果见图 10，我们选取阿拉伯语和僧伽罗语展示结果。图 10：为使模型具备处理广泛爬取 PDF（多语言数据）的能力，我们训练模型具备近 100 种语言的 OCR 能力……

[原文]DF 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 10 , where we select Arabic and Sinhala languages to demonstrate results. Figure 10 : To endow the capability of processing widely crawled PDFs (multilingual data), we train our model with OCR capabilities for nearly 100 languages...

4.1 Vision-text Compression Study

OmniDocBench 上不同文档类别的编辑距离。结果表明，某些文档类型仅需 64 或 100 个 vision token 即可取得良好性能，而其他类型则需要 Gundam 模式。Mode Type 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 0.32 Small 0.085 0.111 0.079 0.147 0.171 0.107 0.131 0.187 0.744 0.205 Base 0.037 0.08 0.027 0.1 0.13 0.073 0.052 0.176 0.645 0.156 Large 0.038 0.108 0.022 0.084 0.109 0.06 0.053 0.155 0.353 0.117 Gundam 0.035 0.085 0.289 0.095 0.094 0……

[原文]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. Mode Type 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 0.32 Small 0.085 0.111 0.079 0.147 0.171 0.107 0.131 0.187 0.744 0.205 Base 0.037 0.08 0.027 0.1 0.13 0.073 0.052 0.176 0.645 0.156 Large 0.038 0.108 0.022 0.084 0.109 0.06 0.053 0.155 0.353 0.117 Gundam 0.035 0.085 0.289 0.095 0.094 0...

4.2 OCR Practical Performance

DeepSeek-OCR 不仅是实验模型，还具有强大的实用能力，可为 LLM/VLM 预训练构建数据。为量化 OCR 性能，我们在 OmniDocBench [27] 上测试 DeepSeek-OCR，结果见表 3。仅需 100 个 vision token（640×640 分辨率），DeepSeek-OCR 即超越使用 256 个 token 的 GOT-OCR2.0 [38]；使用 400 个 token（285 个有效 token，1280×1280 分辨率）时，在该基准上达到与 state-of-the-art 相当的性能。使用少于 800 个 token（Gundam 模式），DeepSeek-OCR 超越 MinerU2.0 [34]，后者 n……

[原文]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 × \times 640 resolution), DeepSeek-OCR surpasses GOT-OCR2.0 [ 38 ] which uses 256 tokens; with 400 tokens (285 valid tokens, 1280 × \times 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 n...

4.3 Qualitative Study

4.3.1 深度解析 DeepSeek-OCR 同时具备 layout 和 OCR 2.0 能力，可通过二次模型调用进一步解析文档内图像，我们将此功能称为「深度解析」。如图 6、7、8、9 所示，模型可对图表、几何、化学公式甚至自然图像进行深度解析，仅需统一 prompt。图 6：在金融研究报告领域，DeepSeek-OCR 的深度解析模式可用于获取文档内图表的结构化结果。图表是金融领域中关键的数据表示形式……

4.3 Qualitative Study

DeepSeek-OCR不仅是一个实验模型；它具有强大的实际能力，可以为LLM/VLM预训练构建数据。为量化OCR性能，我们在OmniDocBench上测试了DeepSeek-OCR。仅需100个视觉token（640×640分辨率），DeepSeek-OCR就超过了使用256个token的GOT-OCR2.0。

[原文]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 11 . Figure 11 : 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, DeepSe...

5 Discussion

我们的工作是对视觉-文本压缩边界的初步探索，研究解码 N 个 text token 需要多少 vision token。初步结果令人鼓舞：DeepSeek-OCR 在约 10× 压缩比下实现近无损 OCR 压缩，20× 压缩仍保留约 60% 准确率。这些发现为未来应用指明方向，例如对多轮对话中超过 k 轮的对话历史实施光学处理以实现 10× 压缩效率。对于更早的 context……

[原文]Our work represents an initial exploration into the boundaries of vision-text compression, investigating how many vision tokens are required to decode N 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 k rounds in multi-turn conversations to achieve 10 × \times compression efficiency. For older contexts,...

6 Conclusion

在本技术报告中，我们提出 DeepSeek-OCR，并通过该模型初步验证了 contexts optical compression 的可行性，表明模型可从少量 vision token 有效解码超过 10 倍数量的 text token。我们相信这一发现将有助于未来 VLM 和 LLM 的发展。此外，DeepSeek-OCR 是高度实用的模型，能够大规模生产预训练数据，是 LLM 不可或缺的助手。当然，仅靠 OCR 不足以完全验证真正的 context optical co……

[原文]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 co...

← 返回首页详细解读