Qwen3.5-9B-Furry-RP-Preview — GGUF Q4_K_M

This is a GGUF Q4_K_M quantized version of the split-MLP slerp merged model.

5.67 GB — ready for llama.cpp, koboldcpp, oobabooga, and any GGUF-compatible backend.

Source model

The source model is a split-MLP slerp merge — not a LoRA adapter, not a full fine-tune. It was created by training a 4B Qwen3.5 on furry roleplay data, then migrating the learned FFN patterns to the 9B base via spherical linear interpolation, with a brief NF4+LoRA polish.

Total training compute: ~2.5 hours on a single RTX 4090.

Why this is different from a normal LoRA

	Normal LoRA	This model
What changes	Low-rank adapters added to existing weights	FFN intermediate dimension expanded via slerp of trained 4B weights
Architecture	Original architecture preserved	`intermediate_size` grows from 12288 → 12416 (128 slerp dims per layer)
Knowledge injection	Rank-limited bottleneck (e.g. rank 16)	Full intermediate dimension weights are updated
Inference	Requires loading base + LoRA adapter	Single standalone model file
Serving	LoRA stacking/merging complexity	Drop-in replacement for original 9B

Usage with llama.cpp

llama-cli \
  -m Qwen3.5-9B-Furry-RP-Preview-Q4_K_M.gguf \
  -p "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n" \
  -n 256 -t 8

Limitations

The base Qwen3.5-9B model has not been abliterated — may refuse or water down NSFW roleplay.
500 quick-adapt steps only — preview release.
FFN down_proj layers fall back to q8_0/q5_0 in this GGUF (12416 dims non-divisible by 256).