Back to today’s briefing

Story perspectives

DeepSeek Launches OCR Model for 20x Text Compression

10/21/2025

44 7 Full Breakdown

Summary

Story summary

DeepSeek has launched DeepSeek-OCR, an open-source model for efficient text compression using visual perception.
The model processes large documents with fewer tokens, achieving up to 20x compression while preserving accuracy.
It generates over 200,000 pages of training data daily on a single A100-40G GPU.
The release includes two main components, DeepEncoder, which compresses images into visual tokens, and DeepSeek3B-MoE-A570M, the decoder.

1 of 1

Story summary

DeepSeek has launched DeepSeek-OCR, an open-source model for efficient text compression using visual perception.
The model processes large documents with fewer tokens, achieving up to 20x compression while preserving accuracy.
It generates over 200,000 pages of training data daily on a single A100-40G GPU.
The release includes two main components, DeepEncoder, which compresses images into visual tokens, and DeepSeek3B-MoE-A570M, the decoder.