license: apache-2.0
language:
- en
- zh
base_model: - Qwen/Qwen3.5-9B
- HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive
base_model_relation: merge
pipeline_tag: image-text-to-text
library_name: transformers
tags: - qwen3_5
- conversational
- qwen
- uncensored
- unfiltered
- vllm
- multimodal
- ablation
- qwen35
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.weightandlm_head.weightare bit-identical. - MLP layers: all
gate_proj,up_proj,down_projweights checked are bit-identical.