DeepSeek-OCR

The DeepSeek-OCR model is a unique multimodal visual-language transformer with 570 million active parameters during inference, designed for efficient optical compression of long text contexts into visual tokens. The key innovation of DeepSeek-OCR lies in the understanding that an image containing document text can represent information with significantly fewer tokens than equivalent digital text. Architecturally, DeepSeek-OCR consists of two main components: DeepEncoder and DeepSeek3B-MoE decoder. DeepEncoder processes images, creating a compressed visual representation of text. The DeepSeek-OCR decoder (based on DeepSeek VL2) reconstructs the original text and structured information from visual tokens. This novel approach allows the model to maintain higher quality than larger models despite its compact size and minimal computational overhead, even when using full attention.

DeepSeek-OCR stands out favorably from other state-of-the-art multimodal models by achieving the required OCR quality with 2-10 times fewer tokens, significantly accelerating and simplifying the processing of voluminous text documents or streams of similar documents. In benchmarks, DeepSeek-OCR demonstrates outstanding results. On the Fox 21 benchmark, it achieves decoding accuracy of approximately 97% with text compression into visual tokens at a ratio of 10, surpassing many contemporary OCR and OCR+visual-text models. On OmniDocBench, DeepSeek-OCR occupies leading positions, using only about 100 tokens for images at 640×640 resolution while maintaining recognition and parsing accuracy for complex structures: formulas, tables, charts, etc. For some document categories (e.g., presentations), the model requires fewer than 64 visual tokens for high-quality recognition.

The model is adaptive and supports multiple operating modes (Tiny, Small, Base, Large, Gundam) for different document types. It is ideally suited for large-scale digitization projects of scanned textual information, recognizing multilingual PDFs (with support for about 100 languages), as well as rendering and structural parsing of documents with tables, formulas, charts, and natural images. Developers recommend DeepSeek-OCR for working with historical archives, documents with long contexts, and automating financial processes.

Announce Date: 20.10.2025
Parameters: 3B
Experts: 64
Activated at inference: 0.57B
Context: 9K
Layers: 12
Attention Type: Full Attention
VRAM requirements: 5.2 GB using 4 bits quantization
Developer: DeepSeek
Transformers Version: 4.46.3
License: MIT

Use our pre-built public endpoints for free to test inference and explore DeepSeek-OCR capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.