nvidia/Mistral-Medium-3.5-128B-NVFP4
LoRA-style vs mistralai/Mistral-Medium-3.5-128B. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of k_proj, o_proj, q_proj, up_proj, v_proj changed, in 72 of 75 layers; norms, biases and embeddings are untouched.
Tensors changed: Embeddings 0/1, Norms 0/248, MLP 4/381, Attention 142/780, other 0/4
1 other repos hold byte-for-byte identical weight files.