llama.cpp asked to take Kimi-K3 vision projector upstream

Kimi-K3 text generation already runs on mainline llama.cpp, but the model is natively multimodal and that path still breaks. Multimodal projection files distributed by Unsloth declare projector type kimik3 (MoonViT-3d: 896-pixel images, patch size 14, 27 blocks, projection dimension 7168, scale factor 2, and 3-D position embeddings). The current CLIP loader rejects them with a failed load.

Daniel Han Chen implemented the missing kimik3 projector in Unsloth's llama.cpp fork. The work is purely additive, wires the new type into the multimodal stack, and has been reported to run vision end-to-end on current master. Maintainers of ggml-org/llama.cpp have been asked to merge the same support so users can load Kimi-K3 image understanding without staying on a fork.