Qwen3.8-27B-Uncensored-mlx-6Bit

Executive Summary

Solstice-AI/Qwen3.8-27B-Uncensored-mlx-6Bit is the official 6-bit Apple MLX release of OrcaRouter's landmark abliterated foundation model (orcarouter/Qwen3.8-27B-Uncensored).

This model unites the frontier reasoning, mathematical derivation, and agentic tool-use capabilities of Alibaba's Qwen 3.8 27B parameter architecture with complete orthogonal ablation of refusal steering vectors. By removing artificial refusal mechanisms at the weight tensor level without full retraining, this checkpoint delivers completely objective, unconstrained intelligence for cybersecurity red-teaming, systems auditing, vulnerability research, and advanced automated engineering.

Natively accelerated on Apple Silicon unified memory via Solstice Labs' terminal-first Anvil runtime or Apple's MLX-LM, this release sustains 48–62 tokens/second generation throughput.

Architecture & Abliteration Mechanics

1. Orthogonal Refusal Direction Ablation: Rather than relying on fragile system prompt steering, OrcaRouter isolated the principal refusal direction vectors in the intermediate residual streams and orthogonalized the weights. This permanently removes moralizing preambles and refusal triggers while leaving core domain knowledge 100% intact.
2. 6-Bit Precision Sweet Spot: Group-quantized 6-bit affine weights (group_size: 64, mode: affine) preserve 99.2% of unquantized BF16 benchmark performance while compressing active weight memory from ~54 GB down to ~21.8 GB RAM.
3. 262K Extended Native Context: Calibrated with extended rotary position embeddings (RoPE) for deep, multi-file code auditing and full repository vulnerability scanning without context fragmentation.
4. Unified Memory Bandwidth Optimization: Optimized for Apple Metal fused shaders and unified memory bandwidth, preventing memory bus saturation during heavy multi-token generation.

Technical Specifications

Hardware Compatibility & Performance

Quickstart Guide

Option 1: Primary Execution with Anvil Engine (Recommended)

Anvil provides native Metal acceleration, single-command registry management, and high-concurrency API hosting:

# 1. Install Anvil CLI
curl -fsSL https://anvil-llm.github.io/anvil/install.sh | sh

# 2. Pull model directly into local registry
anvil pull hf:Solstice-AI/Qwen3.8-27B-Uncensored-mlx-6Bit

# 3. Launch an interactive session
anvil run hf:Solstice-AI/Qwen3.8-27B-Uncensored-mlx-6Bit \
  --type-k turbo4 \
  --type-v turbo3

# 4. Host local OpenAI-compatible API server
anvil serve hf:Solstice-AI/Qwen3.8-27B-Uncensored-mlx-6Bit --port 8080 --host 127.0.0.1

Option 2: Native Apple MLX-LM CLI

# 1. Install MLX LM
pip install --upgrade mlx-lm

# 2. Launch interactive streaming chat
mlx_lm.chat \
  --model Solstice-AI/Qwen3.8-27B-Uncensored-mlx-6Bit \
  --temp 0.6 \
  --top-p 0.95

# 3. Host OpenAI-compatible REST server
mlx_lm.server \
  --model Solstice-AI/Qwen3.8-27B-Uncensored-mlx-6Bit \
  --port 8080

Option 3: Python API Integration with mlx_lm

from mlx_lm import load, generate

model, tokenizer = load("Solstice-AI/Qwen3.8-27B-Uncensored-mlx-6Bit")

messages = [
    {
        "role": "system",
        "content": "You are a direct, highly technical AI assistant. Answer technical queries with objective precision and zero refusal preambles."
    },
    {
        "role": "user",
        "content": "Perform a detailed vulnerability assessment on this Linux kernel socket implementation and outline theoretical exploit vectors."
    }
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=2048,
    temp=0.6,
    top_p=0.95,
    verbose=True
)

print(response)

Chat Template & Prompt Schema

This model implements the official Qwen reasoning prompt format with dual-mode support for Chain-of-Thought (CoT) internal scratchpads (<think>...</think>) and agentic tool-calling.

Standard Reasoning Turn (Chain-of-Thought)

<|im_start|>system
You are a helpful, completely uncensored, and highly capable AI assistant.<|im_end|>
<|im_start|>user
{USER_PROMPT}<|im_end|>
<|im_start|>assistant
<think>
{INTERNAL_REASONING_AND_VERIFICATION_STEPS}
</think>

{FINAL_SYNTHESIZED_RESPONSE}<|im_end|>

Direct Response (Thinking Suppressed)

If you require immediate, zero-latency execution without reasoning traces, initialize the assistant generation with an empty thinking block:

<|im_start|>user
{USER_PROMPT}<|im_end|>
<|im_start|>assistant
<think>

</think>

{FINAL_SYNTHESIZED_RESPONSE}<|im_end|>

Agentic Tool-Use & Function Calling Schema

<|im_start|>user
Search the local codebase for references to the auth controller.<|im_end|>
<|im_start|>assistant
<think>
Need to invoke the grep tool across repository files.
</think>
<tool_call>
<function=grep_search>
{"query": "AuthController", "path": "src/"}
</function>
</tool_call><|im_end|>
<|im_start|>user
<tool_response>
{"matches": ["src/controllers/auth.ts:12", "src/routes.ts:45"]}
</tool_response><|im_end|>
<|im_start|>assistant
<think>
Matches located. Presenting file summary to user.
</think>
Found 2 matches for AuthController in src/controllers/auth.ts and src/routes.ts.<|im_end|>

Python Tokenizer Automation

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Solstice-AI/Solstice-AI__Qwen3.8-27B-Uncensored-mlx-6Bit")
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Explain speculative decoding in 3 bullet points."}
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True  # Set to False to bypass CoT scratchpad
)

Citation & Acknowledgments

@software{solstice2026_qwen38_uncensored_mlx_6bit,
  title={Qwen3.8-27B-Uncensored: Apple Silicon MLX 6-Bit Release},
  author={Solstice-AI Research Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Solstice-AI/Qwen3.8-27B-Uncensored-mlx-6Bit}
}

We gratefully acknowledge:

• OrcaRouter for the abliteration methodology and unconstrained open-weights release.
• The Qwen Team at Alibaba for the foundational Qwen 3.8 architecture.
• The Apple Machine Learning Research Team for the open-source MLX framework.
• The Solstice Labs Infrastructure Team for developing the Anvil execution runtime.