DeepSeek has just unveiled an OCR monster 🤯 DeepSeek-OCR is a...
Charly Wargnier ♨️@DataChaz
100 views
Nov 03, 2025
~1 min read
Advertisement
1
DeepSeek has just unveiled an OCR monster 🤯
DeepSeek-OCR is a 3B-parameter model that redefines document intelligence. It reaches 97% character-level accuracy with 10× input compression, preserving every detail.
Most OCR systems require over 6,000 tokens per page.
DeepSeek-OCR achieves the same with about 100 tokens.
No fidelity loss.
Adaptive modes:
✓ Tiny – simple layouts (≈64 tokens)
✓ Small – standard reports (≈100 tokens)
✓ Database – structured or multi-column docs (≈256 tokens)
✓ Gundam – high-complexity pages (n×100 + 256 tokens)
It scales effortlessly: 200K pages/day on a single NVIDIA A100 GPU.
Its dual-stage design combines a vision encoder for optical compression with a Mixture-of-Experts decoder for structured reconstruction. The outcome is near-perfect precision and zero information loss.
In practice:
➤ 10× lower token costs
➤ Runs on standard hardware
➤ Outputs clean HTML, SMILES, or JSON
While traditional OCRs focus on raw pixel-to-text conversion, DeepSeek-OCR just introduces intelligent semantic parsing 🔥
Open source and free to use.
Repo in 🧵↓
DeepSeek-OCR is a 3B-parameter model that redefines document intelligence. It reaches 97% character-level accuracy with 10× input compression, preserving every detail.
Most OCR systems require over 6,000 tokens per page.
DeepSeek-OCR achieves the same with about 100 tokens.
No fidelity loss.
Adaptive modes:
✓ Tiny – simple layouts (≈64 tokens)
✓ Small – standard reports (≈100 tokens)
✓ Database – structured or multi-column docs (≈256 tokens)
✓ Gundam – high-complexity pages (n×100 + 256 tokens)
It scales effortlessly: 200K pages/day on a single NVIDIA A100 GPU.
Its dual-stage design combines a vision encoder for optical compression with a Mixture-of-Experts decoder for structured reconstruction. The outcome is near-perfect precision and zero information loss.
In practice:
➤ 10× lower token costs
➤ Runs on standard hardware
➤ Outputs clean HTML, SMILES, or JSON
While traditional OCRs focus on raw pixel-to-text conversion, DeepSeek-OCR just introduces intelligent semantic parsing 🔥
Open source and free to use.
Repo in 🧵↓
2
@deepseek_ai Here's the repo:
github.com/deepseek-ai/De…
github.com/deepseek-ai/De…
