4B参数”无审查”模型推荐（基于Qwen系列abliteration改造）

• URL: https://www.threads.com/@weststreetmorning/post/Dc-bwSqIBUA
• Date Saved: 2026-09-07
• Source: Threads (@weststreetmorning)
• Tags: local-llm
• Repo: https://huggingface.co/HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive

Summary

Roundup of best uncensored 4B-parameter models, mostly based on Qwen series with “abliteration” (de-alignment) techniques.

Top picks:

• 🏆 Qwen3-4B / Qwen3.5-4B (HauhauCS Aggressive) — best overall, aggressive de-censoring, 262k context, MMLU ~99.7%
• ⚖️ Heretic variant (Qwen3-4B) — more refined technique, better preserves original capabilities
• 🎭 JOSIE (Gabliterated) — full fine-tune on Qwen3-4B, personality-focused, natural conversation style for roleplay/companionship
• 📱 Qwen3.5-4B-uncensored-MNN — mobile-optimized, MNN format 4-bit quant, ~2.5GB, 18-20 tok/s on flagship phones
• 🧪 Gemma-3-4B-IT-Uncensored-v2 — Google Gemma 3 based, fast response, good instruction following

Other notable:

• Huihui-Qwen3.5-4B-abliterated — popular community base model
• Aisha-Qwen-Uncensored — Qwen3 in GGUF format for local inference
• electroglyph/Qwen3-4B-Instruct-2507-uncensored

Selection tips:

• Check benchmarks (Aggressive hits 99.7% MMLU)
• Ensure format compatibility (GGUF/GPTQ) with your inference framework (Ollama/llama.cpp)
• 4-bit quantization is good balance of performance vs size