DeepSeek has released a new generation optical character recognition system, achieving a technological breakthrough in visual recognition by enabling AI to understand images in a human-like logical sequence. This advancement could reshape application scenarios that rely on complex visual understanding, such as document processing and chart analysis.
On the 27th, DeepSeek launched the DeepSeek-OCR 2 system. This system employs a new method called DeepEncoder V2, allowing AI to &#34;see&#34; images in a logical sequence like humans. The core innovation of this technology lies in changing the traditional way AI processes images. DeepEncoder V2 enables AI to dynamically rearrange image segments based on the meaning of the image, rather than the traditional rigid left-to-right scanning. This approach mimics the way humans follow the logical flow of scenes.
According to the technical report released by DeepSeek, DeepSeek-OCR 2 demonstrates significant advantages across multiple key metrics. In the OmniDocBench v1.5 benchmark test, the model achieved a score of 91.09%, an improvement of 3.73% compared to the previous generation DeepSeek-OCR.
Notably, while maintaining extremely high accuracy, the model strictly controls computational costs, with the number of visual tokens limited to between 256 and 1120, aligning with the upper limit of Google's Gemini-3 Pro. In practical production environments, the model's redundancy rate decreased by 2.08% and 0.81% when processing online user logs and PDF pre-training data, respectively, demonstrating high practical maturity.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/4d614c11-a1f3-4e8e-b7ce-88abbbda2446.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/4d614c11-a1f3-4e8e-b7ce-88abbbda2446.png"/>
<h2>Simulating Human Visual &#34;Causal Flow&#34; Logic</h2>
According to the technical report released by DeepSeek, existing visual language models (VLMs) typically process image slices using a fixed raster scanning order, mechanically scanning from the top left corner to the bottom right corner. The DeepSeek team points out that this method introduces unnecessary inductive bias, contrary to human visual perception. When humans read complex documents, tables, or track spiral lines, their gaze is driven by semantic understanding in a &#34;causal flow,&#34; where the subsequent gaze often causally depends on the previous one, rather than merely moving through spatial coordinates.
Inspired by this cognitive mechanism, the core component of DeepSeek-OCR 2, DeepEncoder V2, is designed to endow the encoder with causal reasoning capabilities. By introducing learnable &#34;Causal Flow Queries,&#34; the model can intelligently reorder visual information during the encoding phase before entering the LLM decoder for content interpretation. This effectively constructs a two-level cascading 1D causal reasoning structure: first, the encoder semantically reorganizes visual tokens, and then the decoder performs autoregressive reasoning on the ordered sequence. This design not only aligns with the nonlinear layout characteristics of optical text, tables, and formulas but also effectively bridges the gap between 2D image structures and 1D language modeling
<h2>Abandoning the CLIP Architecture, Shifting to LLM-style Encoder</h2>
DeepEncoder V2 has implemented significant architectural changes by replacing the original CLIP components in DeepEncoder with a compact LLM-style architecture (specifically Qwen2-0.5B). To achieve parallel processing, the new architecture introduces a set of learnable query vectors, referred to as &#34;causal flow Tokens,&#34; and concatenates the original visual Tokens as a prefix to the sequence.
This architecture adopts a customized attention mask strategy:
<blockquote>
<ul>
<li>Visual Token part: Retains a bidirectional attention mechanism, ensuring that the model can have a global receptive field like CLIP, capturing the overall features of the image.</li>
<li>Causal flow Token part: Utilizes a causal attention mechanism (similar to Decoder-only LLM), where each query Token can only attend to previous Tokens.</li>
</ul>
</blockquote>
Through this design, visual Tokens maintain global interaction of information, while causal flow Tokens gain the ability to reorder visual information. DeepSeek-OCR 2 employs a multi-crop strategy, with the total number of reordered visual Tokens input to the LLM ranging from 256 to 1120, depending on the image resolution. This magnitude is significantly lower than some competitors, which consume over 6000 Tokens, greatly reducing computational overhead while ensuring high performance.
<h2>Significant Performance Improvement and Production Environment Validation</h2>
In the comprehensive evaluation of OmniDocBench v1.5, DeepSeek-OCR 2 performed excellently. Data shows that under the same training data source, the new model achieved a 3.73% performance improvement compared to the DeepSeek-OCR baseline model. Particularly in the edit distance metric for reading order, DeepSeek-OCR 2 significantly reduced from 0.085 to 0.057, directly validating the effectiveness of DeepEncoder V2 in logical reordering.
In addition to benchmark testing, DeepSeek has disclosed the model's performance in actual production pipelines. DeepSeek-OCR 2 primarily serves the online OCR services of DeepSeek-LLMs and PDF pre-training data processing. In a production environment without ground truth, the repetition rate is a core metric for measuring quality. Data shows that when processing online user log images, DeepSeek-OCR 2 reduced the repetition rate from 6.25% to 4.17%; in PDF data production, the repetition rate decreased from 3.69% to 2.88%. This indicates that the new model has extremely high practical value in generating high-quality, low-redundancy text data
<h2>Path to Native Multimodality and True 2D Reasoning</h2>
The release of DeepSeek-OCR 2 is not only an upgrade in OCR performance but also has profound architectural exploration significance. DeepEncoder V2 preliminarily verifies the potential of using language model architecture as a visual encoder. This architecture naturally inherits the achievements of the LLM community in infrastructure optimization, such as the mixture of experts (MoE) architecture and efficient attention mechanisms.
The DeepSeek team believes this provides a promising path toward a unified multimodal encoder. In the future, a single encoder may achieve feature extraction and compression of images, audio, and text within the same parameter space by configuring learnable queries specific to each modality. The &#34;two cascaded 1D causal reasoners&#34; model demonstrated by DeepSeek-OCR 2, which decomposes 2D understanding into two complementary sub-tasks of &#34;reading logical reasoning&#34; and &#34;visual task reasoning,&#34; may represent a groundbreaking architectural approach to achieving true 2D reasoning.
More news, ongoing updates

DeepSeek-OCR 2 Heavy Release: AI Learns "Human Visual Logic" to Interpret Images through Causal Flow