DeepSeek drops new model DeepSeek-OCR

What happened: On October 20, DeepSeek dropped a new model – DeepSeek-OCR.

Some context: OCR, or optical character recognition, is a technology that can read text from images – a feature already common in smartphones.

DeepSeek-OCR does that too, but its real innovation is in its compression technique. 

• DeepSeek-OCR compresses text images (like scanned documents) into visual representations that LLMs can read. 
• This approach is more computationally efficient than feeding LLMs pure text. 

How good is the compression? DeepSeek claims its approach compresses data by about 10× while maintaining 97% accuracy.

Why it matters: LLMs face significant computational bottlenecks when processing long streams of text, a growing problem as AI agents handle more complex tasks.

• DeepSeek’s novel approach could help alleviate those challenges – and unlock AI agents’ ability to do more useful things with less computing power.

The bottom line: DeepSeek’s much-hyped R2 model may be delayed, but the lab keeps dropping impressive innovations.