DeepSeek-OCR

Name: DeepSeek-OCR
Author: DeepSeek

DeepSeek · released 2025-10-17 · data checked 2026-11-06

DeepSeek-OCR is a 3.3B-parameter OCR and document-parsing model from DeepSeek (CN), released 2025-10-17 under the MIT license. It scores 75.7 on olmOCR-Bench (#18 of 24). It needs about 7.7 GB of memory at 16-bit.

The original "optical compression" model; succeeded by DeepSeek-OCR 2.

OmniDocBench v1.6

—

Not reported

olmOCR-Bench

75.7

#18 of 24 models

Parameters

3.3B

570M active · Mixture-of-Experts

Memory to run

~7.7 GB

16-bit weights plus about 1 GB

Benchmarks

olmOCR-Bench Leaderboard75.7

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: 3.3B
Active parameters: 570M
License: MIT
Memory to run: ~7.7 GB at 16-bit
Released: 2025-10-17

Weights and usage on Hugging Face ↗

Closest alternatives

MonkeyOCR-pro 3B (3.8B, 75.8 on olmOCR-Bench, Non-commercial)
GLM-OCR (1.3B, 75.2 on olmOCR-Bench, MIT)
MinerU2.5 (2509) (1.2B, 75.2 on olmOCR-Bench, AGPL-3.0)
DeepSeek-OCR 2 (3.4B, 76.3 on olmOCR-Bench, Apache-2.0)

Sources

https://huggingface.co/deepseek-ai/DeepSeek-OCR

Something wrong here? Report it.