DarkRomeo88/Qwen3.8-27B-uncensored-heretic-w4a8-ComfyUI
LoRA-style vs Qwen/Qwen3.8-27B. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of conv1d, o_proj changed, in 8 of 64 layers; norms, biases and embeddings are untouched.
Tensors changed: Norms 0/325, Attention 8/359, MLP 0/121, Embeddings 0/5, lm_head 0/1, other 0/1
2 other repos hold byte-for-byte identical weight files.