Chinese AI startup DeepSeek has released an upgraded OCR model, DeepSeek-OCR 2, which incorporates Alibaba Cloud’s open-source Qwen2-0.5b model to improve performance, according to a research paper.

The new model replaces a component of its original architecture that used contrastive language image pre-training (CLIP), a neural network framework developed by OpenAI.

DeepSeek claims that this change allows its OCR system to process documents in a manner similar to human reading, by following flexible and semantically coherent scanning patterns.

Benchmark tests indicate a 3.7% performance increase over the previous version.

The company has open-sourced DeepSeek-OCR 2 on Hugging Face, a platform for AI development.

🔗 Source: South China Morning Post

Recent DeepSeek developments