A third-party test model that bolts Kimi's vision encoder onto DeepSeek V4 Flash.
A bookworm robot finally gets eyes. But the eyes are borrowed from Kimi next door. Neither brain was retrained. Someone just soldered a small plug in between.
It can try simple image tasks. It is not proven for big real-world use yet.
DeepSeek
This project is built on the DeepSeek V4 Flash text backbone.
Kimi
Its eyes are borrowed: MoonViT, the vision encoder from Kimi K2.6.
VLM
The two backbones together work as a test VLM.
Fine-tuning
It freezes both backbones and trains only their small connecting layer.