dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8
LoRA-style vs deepseek-ai/DeepSeek-V4.1-Flash. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of w2, wo_b changed, in 27 of 40 layers; norms, biases and embeddings are untouched.
Tensors changed: other 0/17, Attention 27/542, MLP 27/1149, Norms 0/248, Embeddings 0/6
2 derivatives checked
- pyrodog/DeepSeek-V4.1-Flash-UNCENSORED-DwarfStar-Q2: Same weights, repacked (fewer layers)
- MomonV/DeepSeek-V4.1-Flash-UNCENSORED-NVFP4: Quantization
25 other repos hold byte-for-byte identical weight files.