nvidia/Phi-4-multimodal-instruct-FP8
Adapter vs microsoft/Phi-4-multimodal-instruct. nvidia/Phi-4-multimodal-instruct-FP8 is a LoRA adapter, not full weights: rank 256-320, on down_proj, gate_up_proj, o_proj, qkv_proj in 32 layers (256 adapted matrices). It modifies microsoft/Phi-4-multimodal-instruct at load time.