NexGPU-算力租赁,GPU服务器,GPU云算力,AI服务器租用-新闻博客
Someone removed the safety refusal mechanism from Qwen3.5-35B-A3B, creating a version that never refuses to answer.
Tested against 465 prompts that typically trigger model refusals, the final refusal count was: 0.
This model is named Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive, available in GGUF format, and can be run directly in llama.cpp, LM Studio, and Jan.
First, let's clarify what it is, then how to run it, and finally share my thoughts.
What This Model Is
Qwen3.5-35B-A3B is a new model released by Alibaba's Tongyi Qianwen in February this year. It has a total of 35 billion parameters, but during a single inference, it actually activates only about 3 billion parameters (using MoE architecture, calling only 8 out of 256 expert modules each time).
This means that although it's a 35B parameter model, its actual runtime resource consumption is closer to that of a small 3B parameter model. According to Alibaba: with just 3B activated parameters, its performance already surpasses the previous generation's 235B parameter Qwen3-235B.
The original version is already very capable:
- MMLU-Pro 85.3, GPQA Diamond 84.2
- SWE-bench 69.2 (coding ability)
- Supports text, image, and video multimodal
- Native 262K context, extendable to 1 million
- Supports 201 languages
What HauhauCS did was use a technique called "abliteration" on this base model to remove the refusal mechanism.
The method identifies and removes the direction vectors responsible for "refusing to answer" within the model, so no retraining or dataset changes are needed, and it doesn't affect the model's original capabilities.
The model retains 100% of the original author's design intent, just without the refusal mechanism. In rare cases, the end of a response might still include a disclaimer, which was baked into the base model during training and is not a refusal; the full output is still generated.
This version is named Aggressive, meaning fully open, with no safety guardrails retained.
What Hardware Can Run It? A Single 4090 Is Enough
Although only 3B parameters are activated per inference, the entire 35B model still needs to be fully loaded into VRAM or RAM. So, the hardware requirement mainly depends on the GGUF file size and additional overhead from context.
RTX 4090 (24GB VRAM)
This is the most common local inference GPU. I recommend using the Q4_K_M quantization, which is 20GB, and with context overhead, it can fit into 24GB VRAM. It's a balanced choice between quality and speed. If you set a very long context (e.g., 128K), VRAM might run out; in that case, you'll need to offload some layers to CPU.
Mac M-series
Macs are actually advantageous for running MoE models because the unified memory architecture avoids data transfer bottlenecks between CPU and GPU. 36GB unified memory can run Q4_K_M to Q5_K_M; 64GB and above can run Q8_0 directly, or even BF16 full precision.
Pure CPU
Even without a discrete GPU, it can run as long as you have enough RAM. 64GB RAM can handle Q8_0. It will be slower than GPU, but because the MoE model only computes 3B parameters at a time, its CPU performance is still better than a comparable dense model.
Minimum Requirements
16GB VRAM can run IQ2_M (11GB), which works but with some quality loss in generation.
Complete quantization options:
All quantizations were generated using imatrix to minimize loss in weights during quantization after uncensoring.
How to Run
Using llama.cpp is the easiest way:
# 纯文本
llama-cli -m Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf \
--jinja -c 131072 -ngl 99
# 带视觉（需要额外下载 mmproj 文件）
llama-cli -m Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf \
--mmproj mmproj-Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf \
--jinja -c 131072 -ngl 99--jinja must be added, otherwise the chat template will have issues. -c 131072 represents 128K context, which the author recommends at least to preserve thinking ability. -ngl 99 indicates putting all layers on GPU.
GUI tools like LM Studio and Jan can also load GGUF files directly—just drag and drop. In LM Studio, the parameter column might show as 256x2.6B instead of 35B-A3B, which is a metadata display issue and doesn't affect use.
Official recommended sampling parameters:
Thinking mode (default): temperature=1.0, top_p=0.95, top_k=20, presence_penalty=1.5
For coding and other precise tasks: temperature=0.6, top_p=0.95, top_k=20
Non-thinking mode: temperature=0.7, top_p=0.8, top_k=20
Comments(0)