DeepSeek Open-Sources DeepSeek-OCR 2: Visual Causal Flow for Markdown Conversion

tom_doerr · x · 2026-08-11

DeepSeek has officially open-sourced the DeepSeek-OCR 2 model and its codebase. The model introduces "Visual Causal Flow" technology to explore more human-like visual encoding, enabling highly efficient conversion of images and PDF documents into Markdown format.

The project has already garnered 3.2k stars on GitHub. For environment setup, the official repository provides detailed installation guides based on CUDA 11.8 + Torch 2.6.0, supporting inference deployment via both vLLM (version 0.8.5) and Transformers.

Original post →

More from Models

Models channel →