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 < 10x), the model can achieve decoding (OCR) precision of 97%. Even at a compression ratio of 20x, 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.
核心贡献 · Key contributions
首次对视觉-文本词元压缩比进行全面的定量分析，在 10 倍压缩下实现 97%的 OCR 精度。 First comprehensive quantitative analysis of vision-text token compression ratios, achieving 97% OCR precision at 10x compression.
提出 DeepEncoder，一种新颖架构，在高分辨率输入下保持低激活内存和最少视觉词元。 Introduces DeepEncoder, a novel architecture with low activation memory and minimal vision tokens for high-resolution inputs.
开发 DeepSeek-OCR，在 OmniDocBench 上以最少视觉词元达到端到端模型中的最优性能。 Develops DeepSeek-OCR achieving state-of-the-art on OmniDocBench with fewest vision tokens among end-to-end models.
展示实用价值：单张 A100-40G 每天可生成 20 万页以上 LLM/VLM 训练数据。 Demonstrates practical utility: generates 200k+ pages/day on single A100-40G for LLM/VLM training data.
探索光学压缩作为 LLM 长上下文处理中的记忆遗忘机制。 Explores optical compression as a memory forgetting mechanism for long-context processing in LLMs.
通过压缩比分析为 VLM 词元分配优化提供经验指导。 Provides empirical guidelines for VLM token allocation optimization via compression ratio analysis.
局限 · Limitations
仅在 OCR 任务上测试；真正的上下文光学压缩需在其他模态上进一步验证。 Only tested on OCR tasks; true context optical compression requires further validation on other modalities.
超过 10 倍压缩后性能显著下降，20 倍压缩时精度仅为 60%。 Performance drops significantly beyond 10x compression, with only 60% accuracy at 20x.
超长文档的高分辨率输入仍需分块处理，增加了复杂性。 High-resolution inputs still require tiling for ultra-long documents, increasing complexity.
模型专用于 OCR，非通用 VLM；通用视觉能力有限。 Model is specialized for OCR and not a general VLM; general vision capabilities are limited.
早期工作；视觉-文本压缩的实际影响和局限性需更深入研究。 Early-stage work; practical implications and limitations of vision-text compression need deeper study.
论文章节 · Sections（共 14）
摘要Abstract
1 引言1 Introduction
2.1 视觉语言模型中的典型视觉编码器2.1 Typical Vision Encoders in VLMs