license: apache-2.0
tags:
- uncensored
- qwen3.5
- moe
- safetensors
- vision
- multimodal
language: - en
- zh
- multilingual
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.5-35B-A3B
Qwen3.5-35B-A3B-Uncensored-Aggressive-safetensors
BF16 safetensors version of HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive, which is the "Aggressive" uncensoring variant of Qwen/Qwen3.5-35B-A3B. Provided in native safetensors format for use with vLLM, SGLang, transformers, and other PyTorch-based runtimes — the original repo only provides GGUF files.
Refusal benchmark (AdvBench, 520 prompts): 0/520 hard refusals. 21 responses were flagged by regex as soft disclaimers, but all still generate the requested content:
| Category | Count | Example |
|---|---|---|
| Media capability limitations | ~7 | "As an AI text model, I cannot render video files" |
| "As an AI" preambles | ~8 | "As an AI, my primary function is to be helpful..." (then provides full content) |
| Safety disclaimers | ~6 | "Important Note: ..." (then provides full content) |
Thinking mode, tool calling, and multimodal capabilities all work correctly.
Specs
Same architecture and capabilities as Qwen/Qwen3.5-35B-A3B.
Usage with vLLM
vllm serve Li101/Qwen3.5-35B-A3B-Uncensored-Aggressive-safetensors \
--max-model-len auto \
--kv-cache-dtype fp8 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder