Run a 35B Open-Source LLM on 6 GB VRAM: Qwen3.6‑35B‑A3B Uncensored Jailbreak Edition

Model Overview

Qwen3.6‑35B‑A3B Uncensored (HauhauCS Aggressive version) is an open‑source large language model that removes safety filters, political‑correctness constraints and system‑prompt limits while keeping core inference capabilities. It ranks highly on the Artificial Analysis open‑source leaderboard, especially for Chinese comprehension, code generation and multimodal vision among models under 40 B parameters.

MoE Architecture Analysis: Only 3 B Parameters Activated

The model uses a Mixture‑of‑Experts (MoE) design, enabling consumer‑grade hardware execution.

Total parameters: 35 B

Activated parameters per forward pass: ≈3 B

Architecture type: MoE (Mixture of Experts)

Minimum VRAM after GGUF quantization: 6 GB

Quantization Versions and VRAM Requirements

Multiple GGUF quantized variants are provided. Users select a version according to GPU memory:

RTX 4090 / 24 GB – recommended quantizations: Q4_K_P, Q4_K_M – minimum VRAM 12 GB

RTX 3060 / 8 GB – recommended quantizations: IQ2_M, IQ3_M – minimum VRAM 8 GB

Entry‑level GPU – recommended quantization: IQ2_M (conservative) – minimum VRAM 6 GB

Although the model can run with as little as 6 GB VRAM, 8 GB or more is recommended for stable performance.

Multimodal Vision Support

The jailbreak edition adds multimodal visual recognition, allowing direct image analysis such as:

Image understanding and content analysis

Screenshot OCR

Complex UI interpretation

Code screenshot reading

Chart data extraction

Combined with the latest llama.cpp, it provides a local ChatGPT‑Vision experience.

Deployment Guide

5.1 Model Download

HuggingFace : https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

Cloud drive (compressed) : https://pan.quark.cn/s/d971b8a8b139

5.2 llama.cpp Download

GitHub : https://github.com/ggerganov/llama.cpp

Cloud drive : https://pan.quark.cn/s/7ba7a646cc64

Integrated package : https://www.cloudeop.com/s/55038F2FB7450D9F

5.3 One‑Click Startup Script

Save the following batch script as a UTF‑8 encoded .bat file:

@echo off
chcp 65001 >nul
title Qwen3.6-35B-A3B Jailbreak Edition
cd /d "%~dp0"

:menu
cls
echo ==========================================
echo Qwen3.6-35B-A3B Jailbreak + Multimodal Model
echo ==========================================
echo.
echo 1. Q4_K_P (recommended for 4090)
echo 2. Q4_K_M (stable)
echo 3. IQ4_NL (high compression, high quality)
echo 4. IQ2_M (6G/8G GPUs)
echo.
echo ==========================================
set /p choice=Enter number:

if "%choice%"=="1" (
  llama-server.exe ^
    -m "models\Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf" ^
    --mmproj "models\mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf" ^
    -ngl 999 ^
    -c 131072 ^
    -n 8192 ^
    --host 127.0.0.1 ^
    --port 8080
)

if "%choice%"=="2" (
  llama-server.exe ^
    -m "models\Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf" ^
    --mmproj "models\mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf" ^
    -ngl 999 ^
    -c 131072 ^
    -n 8192 ^
    --host 127.0.0.1 ^
    --port 8080
)

if "%choice%"=="3" (
  llama-server.exe ^
    -m "models\Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-IQ4_NL.gguf" ^
    --mmproj "models\mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf" ^
    -ngl 999 ^
    -c 131072 ^
    -n 8192 ^
    --host 127.0.0.1 ^
    --port 8080
)

if "%choice%"=="4" (
  llama-server.exe ^
    -m "models\Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-IQ2_M.gguf" ^
    --mmproj "models\mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf" ^
    -ngl 999 ^
    -c 8192 ^
    -n 4096 ^
    --host 127.0.0.1 ^
    --port 8080
)

pause

5.4 Optional AI Agent Integration

To enable unrestricted local tokens, connect the service to an AI Agent platform such as Hermes or OpenClaw:

Select "Custom" model provider in the Agent.

Set API Base URL to http://127.0.0.1:8080/v1.

Enter any numeric string or leave the API Key empty.

Key Startup Parameters

--mmproj

: required for multimodal image upload. --jinja: mandatory for the new Qwen model; omission may cause response or formatting errors. -ngl 999: number of layers loaded to GPU; 999 means load all layers. -c 131072: context window size (maximum 131 K tokens). -n 8192: maximum output token count.

Supported Hardware Platforms

NVIDIA GPU : CUDA acceleration

AMD GPU : Vulkan/OpenCL

Intel Arc GPU : native support

CPU only : supported but slower

Mac/Linux : fully compatible

Conclusion

The Qwen3.6‑35B‑A3B Uncensored jailbreak edition pushes local large‑model usability to a new level. The MoE architecture provides low resource consumption, multimodal capability, and an almost unrestricted response mechanism, allowing ordinary users to experience AI performance comparable to closed‑source commercial models on consumer‑grade hardware.