DeepSeek’s Hidden Multimodal Model: Technical Deep‑Dive and Unexpected Bugs

Unexpected Bug in Real‑World Testing

When the community queried the model with a photo of Liang Wenfeng, it confidently answered “This is Zhang Yiming,” mistakenly identifying the ByteDance founder as the CEO; similar failures occurred with other faces, while it correctly recognized Sam Altman.

Technical Strengths of DeepSeek‑V4‑Flash‑Vision‑Exp

Despite the quirks, the model’s ability to recognize objects, scenes, and logos is solid, and its visual‑question‑answering performance is respectable. The underlying technique builds on DeepSeek’s previously published visual primitives.

Visual Primitives and Spatial Reasoning

The model treats bounding boxes and coordinate points as minimal thinking units, effectively “pointing” while reasoning. This enables strong results on pure visual tasks such as counting (89.2% accuracy), maze navigation (66.9% – second best 50.6%), and path tracing (56.7% – second best 46.5%). It retains only about 90 visual tokens in the KV cache, delivering efficiency far beyond models like GPT‑5.4.

Model Architecture and Compression

DeepSeek‑V4‑Flash‑Vision‑Exp uses the DeepSeek‑V4‑Flash language backbone (284 B total parameters, 13 B active MoE) and a self‑developed DeepSeek‑ViT visual encoder that accepts arbitrary‑resolution inputs. For a 756×756 image, raw pixels (571,536) are reduced to 2,916 ViT patch tokens, then to 324 tokens after 3×3 spatial compression, and finally to 81 visual KV entries via CSA (Compressed Sparse Attention), achieving a 7,056 : 1 compression ratio without losing key information.

Takeaways

The model excels at complex chart visualization and spatial reasoning, but it still struggles with face recognition. Users should leverage its strengths for visual analysis while being aware of its current limitations.

https://api-docs.deepseek.com/guides/vision/
https://github.com/deepseek-ai/Thinking-with-Visual-Primitives/blob/main/Thinking_with_Visual_Primitives.pdf