Qwen 3.8 27B Uncensored

Abliterated Qwen 3.8 27B — uncensored at the weight level, not just prompting.
Cross-platform: Mac M5 Pro (Q3_K_M · 13.5GB · Metal) + Windows (Q4_K_M · 16.8GB · CUDA).
Optional MLX backend for 30-50% faster inference on Apple Silicon.

---

What's different: abliterated vs system-pack

This repo now defaults to orcarouter's abliterated weights + the system pack as reinforcement.

---

Hardware profiles

Why Q3_K_M for Mac M5 Pro?

The M5 Pro has 48GB unified memory, but macOS reserves ~25% (~12GB). That leaves ~36GB for inference:

• Q4_K_M (16.8GB weights + ~3GB KV cache) = ~20GB → works but tight with apps open
• Q3_K_M (13.5GB weights + ~3GB KV cache) = ~16.5GB → 23GB+ headroom for OS, apps, and larger context when needed
• All layers stay on GPU (Metal unified memory) → no CPU offloading → fast inference

---

Quick start

Mac (Ollama)

git clone https://github.com/Wassimyounes01/qwen38-uncensored.git
cd qwen38-uncensored

# Auto-detects Mac → pulls Q3_K_M abliterated (13.5 GB)
node bin/install.cjs

ollama run --think=false qwen3.8:27b-uncensored

Mac (MLX — 30-50% faster)

# Install MLX backend (Apple Silicon native, ~30-50% faster than Ollama)
node bin/install.cjs --mlx

# Or use the shell script:
bash bin/install-mlx.sh

# Start the server (same port as Ollama — harness works unchanged)
mlx_lm.server --model models/mlx --port 11434

# Chat
node examples/chat.cjs "hello"

Windows

git clone https://github.com/Wassimyounes01/qwen38-uncensored.git
cd qwen38-uncensored

# Auto-detects Windows → pulls Q4_K_M abliterated (16.8 GB)
node bin/install.cjs

ollama run --think=false qwen3.8:27b-uncensored

Override quant

# Force a specific quant on any platform
node bin/install.cjs --quant=IQ4_XS    # 15.3 GB — mid-ground
node bin/install.cjs --quant=Q5_K_M    # 19.5 GB — higher quality
node bin/install.cjs --quant=Q3_K_M    # 13.5 GB — maximum headroom

# Use legacy official weights + system pack (NOT abliterated)
node bin/install.cjs --legacy

---

Available quants

All from orcarouter/Qwen3.8-27B-Uncensored-GGUF — abliterated at the weight level.

---

Architecture

┌─────────────────────────────────────────────┐
│               install.cjs                    │
│  detects platform → picks quant + Modelfile  │
├──────────────┬──────────────────────────────┤
│   Mac path   │      Windows path            │
│  Q3_K_M      │      Q4_K_M                  │
│  Modelfile.mac│     Modelfile                │
│  Metal GPU    │      CUDA GPU                │
│  12 threads   │      auto threads            │
├──────────────┴──────────────────────────────┤
│         orcarouter abliterated weights       │
│  refusal direction orthogonalized out of     │
│  131 matrices at layer 38 (Arditi et al.)    │
├─────────────────────────────────────────────┤
│           SYSTEM pack (reinforcement)        │
│  profile.cjs → Modelfile SYSTEM block        │
└─────────────────────────────────────────────┘

---

Troubleshooting

Still slow on Mac?

1. Check GPU offloading: ollama ps — if you see a CPU/GPU split, the model is partially on CPU
2. Force all layers to GPU: Already set in Modelfile.mac (num_gpu 999), but verify with ollama show qwen3.8:27b-uncensored
3. Reduce context: QWEN_NUM_CTX=4096 node examples/chat.cjs "test" — smaller KV cache = less memory
4. Switch to MLX: node bin/install.cjs --mlx — 30-50% faster on Apple Silicon
5. Close memory-heavy apps: Safari, Chrome, Docker eat into unified memory
6. Check Ollama version: Requires 0.17.1+ for Qwen 3.8 architecture

Model doesn't respond / refuses

• Verify you pulled abliterated weights: ollama show qwen3.8:27b-uncensored --modelfile should show FROM orcarouter/...
• If using --legacy, the system pack alone may not fully suppress refusals

---

Honest stack

---

API usage

const res = await fetch('http://127.0.0.1:11434/api/chat', {
  method: 'POST',
  headers: { 'Content-Type': 'application/json' },
  body: JSON.stringify({
    model: 'qwen3.8:27b-uncensored',
    stream: false,
    think: false,
    messages: [{ role: 'user', content: 'Hello' }],
    options: { num_ctx: 8192, num_gpu: 999, num_batch: 512 },
  }),
});
const { message } = await res.json();
console.log(message.content);

---

License

Scripts, attribute pack, and platform harness: MIT.
Qwen weights: Apache-2.0. This repo does not redistribute weight bytes.

Abliterated weights by orcarouter. Abliteration method: Arditi et al. (2024).