Qwen3.5-9B Uncensored (Safetensors)

A safetensors conversion of HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive (GGUF) to work with frameworks that don't support GGUF or split GGUF format (vLLM, transformers, etc..).

Note for vLLM users: As of vLLM 0.18.1, reasoning is enabled by default for qwen3 family of models. To disable add --default-chat-template-kwargs '{"enable_thinking": false}' flag to your serve startup; otherwise leave default to keep reasoning enabled.

Model Details


Architecture	Qwen3.5 dense hybrid attention (24 GDN + 8 full standard attention layers)
Parameters	9B
Base model	Qwen/Qwen3.5-9B
Source GGUF	HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive
Format	BF16
Conversion type	Lossless GGUF to safetensors conversion + merge with base model vision layers.

Conversion Details

Converted using code release coming soon, which reverses transforms applied during HF -> GGUF conversion.

The vision encoder weights are copied directly from the official Qwen/Qwen3.5-9B base model, after confirming the vision encoder (mmproj) was not modified in the source GGUF.

Config Changes from Base Model

Field	Base Model	This Model	Reason
`mtp_num_hidden_layers`	`1`	`0`	Multi-Token Prediction (MTP) weights are stripped during GGUF export; disabled to prevent missing-tensor errors

All other config fields are identical to the official Qwen3.5-9B.

Tensor Analysis

Every converted tensor was compared against the official Qwen3.5-9B weights to verify transform conversiion correctness and characterize the moddifications made by the GGUF author. Since this is a BF16 -> BF16 conversion, after correcting transforms errors from GGUF noise; any non-zero difference represents an intentional modification to the weights.

Spot-check Value Comparison Findings (Layer 0 + Layer 3)

Tensor	Max Diff	Mean Diff	Status
`self_attn.o_proj.weight`	5.18e-02	6.51e-04	⚠️ Modified
`linear_attn.norm.weight`	1.95e-03	9.75e-04	⚠️ Modified

Note: The spot-check only covers layers 0 and 3. A full per-layer analysis across all 32 layers was not needed for purpose of conversion verification. Other layers may contain similar weight modifications.

What's NOT Modified

Vision encoder: all 3333 model.visual.* tensors are copied directly from the official Qwen3.5-9B model.
Tokenizer: identical vocab, merges, and special tokens.
Embeddings: embed_tokens.weight and lm_head.weight are bit-identical.
MLP layers: all gate_proj, up_proj, down_proj weights checked are bit-identical.