DeepSeek Vision Impresses in Tests but Still Struggles with Spatial Reasoning

teortaxesTex · x · 2026-08-09

User tests reveal that DeepSeek's vision capabilities are highly competitive, even surpassing Claude and GPT-4o in identifying fine details in complex images. However, when it comes to 3D spatial reasoning—such as determining the correct orientation for a radial cross-hole—DeepSeek makes mistakes similar to other mainstream LLMs, highlighting a persistent weakness in spatial awareness.

Original post →

More from Models

Models channel →