Document OCR model compressing page images into few vision tokens (DeepEncoder + 3B MoE decoder).
Released
Oct 20, 2025
Context window
Not disclosed
Max output
Not disclosed
Parameters
3B 0.6B active
Input price
Not disclosed
Output price
Not disclosed
Prices are first-party list prices in USD at the standard tier. Last verified Oct 30, 2026.
Pricing
No first-party API. This is an open-weights release you can run yourself.
Specifications
- Developer
- DeepSeek
- Model type
- Multimodal
- Architecture
- MoE
- Parameters
- 3B (0.6B active)
- Input
- Text Image
- Output
- Text
- Knowledge cutoff
- Not disclosed
- License
- MIT License
- API access
- No first-party API
- Capabilities
- Small LM Vision