tinyanalysis

OCR models / DeepSeek

DeepSeek-OCR 2

DeepSeek · released 2026-01-27 · data checked 2026-11-06

DeepSeek-OCR 2 is a 3.4B-parameter OCR and document-parsing model from DeepSeek (CN), released 2026-01-27 under the Apache-2.0 license. It scores 90.25 on OmniDocBench v1.6 (#26 of 33) and 76.3 on olmOCR-Bench (#16 of 24). It needs about 7.8 GB of memory at 16-bit.

Compresses each page into few vision tokens; 3B Mixture-of-Experts decoder with about 570M active.

OmniDocBench v1.6
90.25
#26 of 33 models
olmOCR-Bench
76.3
#16 of 24 models
Parameters
3.4B
570M active · Mixture-of-Experts
Memory to run
~7.8 GB
16-bit weights plus about 1 GB

Benchmarks

OmniDocBench v1.6 Leaderboard90.25
Best: OXR 1.0 97.03
olmOCR-Bench Leaderboard76.3
Best: OXR 1.0 87.8

The tick marks the best score on each benchmark. See how to read the ranking.

Specifications

Lab
DeepSeek (CN)
Type
OCR / document parsing specialist
Total parameters
3.4B
Active parameters
570M
License
Apache-2.0
Memory to run
~7.8 GB at 16-bit
Released
2026-01-27
Weights and usage on Hugging Face ↗

Closest alternatives

  • HunyuanOCR 1.0 (1B, 89.95 on OmniDocBench, Tencent Hunyuan Community License)
  • dots.ocr (3B, 90.77 on OmniDocBench, MIT)
  • Dolphin v2 (3.8B, 89.50 on OmniDocBench, Not stated)
  • jina-ocr-v1 (3.4B, 91.14 on OmniDocBench, CC-BY-NC-4.0)

Sources

Something wrong here? Report it.