Qwen3.5-9B for ComfyUI

ComfyUI-format LoRAs for the Qwen3.5-9B text encoder.

File	Size
loras/qwen35-9b-hauhaucs_uncensored_lora_rank4_fp32.safetensors	13.4 MB	Reproduces HauhauCS Uncensored Aggressive on the stock encoder instead of a second checkpoint. Removes refusal behaviour.

Use

LoRAs into models/loras/, plus qwen3.5_9b_bf16.safetensors from Comfy-Org/Qwen3.5 into models/text_encoders/.

CLIPLoader -> LoRA on CLIP (strength 1.0) -> Generate Text

• Leave CLIPLoader's type on the default — Qwen3.5-9B is detected from the weights.
• Strength 1.0 is the full uncensored model; lower values blend back toward stock behaviour.
• Use greedy decoding (sampling_mode: off). Qwen3.5 thinks by default; the checks below were run with thinking off.

How it was made

A plain weight diff, uncensored − Qwen/Qwen3.5-9B, factored per matrix with SVD. The upstream edit is exactly rank 4 — every changed matrix has four real singular values, then a cliff to bf16 rounding noise — so rank 4 keeps ≥99.9% of it (rank 1 would keep ~65%).

• 68 matrices, language model only: attention outputs (linear_attn.out_proj / self_attn.o_proj, layers 3–31), mlp.down_proj (12–25), mlp.gate_proj / mlp.up_proj (19–31). Embeddings, lm_head and the vision tower are untouched.
• Dropped: sub-rounding (~1e-3 relative) changes to 1-D norm weights, left over from the GGUF round-trip, which a LoRA can't carry anyway.
• Checked against the uncensored weights in transformers: next-token KL to the uncensored model goes from 0.337 (stock) to 0.0007 (stock + LoRA), and greedy outputs match it on refusal prompts while benign answers are unaffected.

Keys are ComfyUI's generic text-encoder form (text_encoders.transformer.model.layers.N.<module>.lora_{down,up}.weight), fp32, no alpha.

Credits

Qwen3.5-9B (Apache-2.0) · HauhauCS (Apache-2.0) — all uncensoring behaviour is theirs, this only re-expresses it as a LoRA · nDimensional (Apache-2.0) for the safetensors conversion the diff was taken from.