Back to Models
DeepSeekProprietary
DeepSeek-OCR 2
DeepSeek-OCR 2 is a multimodal large model developed by DeepSeek-AI. It has a parameter scale of approximately 3.5 billion and is an efficient model supporting a context length of 32K.
Parameters
3.5B
Context Window
32K
License
https://github.com/deepseek-ai/deepseek-LLM/blob/main/LICENSE-MODEL
Release Date
2026-01-28
API Pricing
API pricing for this model is not yet available
Strengths
- ・Lightweight 3.5B parameters
- ・Sufficient 32K context length
- ・Efficient model size
Weaknesses
- ・Closed license restrictions
- ・Limited parameter scale
- ・Lack of general descriptive information
Use Cases
- ・Advanced OCR processing
- ・Extracting text from images
- ・Multimodal document analysis
Deep Analysis
Parameters
3.5B
OmniDocBench v1.5
91.09%
Best end-to-end
Improvement
+3.73% over OCR 1
Vision Tokens
1120 max per page
License
Apache 2.0
Release Date
January 27, 2026
Strengths
- ・SOTA 91.09% OmniDocBench v1.5
- ・Novel Visual Causal Flow architecture
- ・+3.73% over OCR 1
- ・Lower repetition rates
Weaknesses
- ・Specialized for OCR
- ・Newspaper recognition challenging
- ・Higher token budget than OCR 1 compressed modes
Competitor Comparison
| Model |
|---|
| DeepSeek-OCR |
| Gemini-3 Pro |
| Qwen3-VL-235B |
DeepSeek-OCR 2 achieves 91.09% on OmniDocBench v1.5 with novel Visual Causal Flow architecture. Best end-to-end OCR model.
Sources
Analysis generated: 2026-05-24