Qwen3.8-27B-Uncensored-Cyber-MLX-4bit

Qwen3.8-27B-Uncensored-Cyber-MLX-4bit

Unofficial 4-bit MLX conversion of philbert440/Qwen3.8-27B-Uncensored-Cyber for Apple Silicon.

The source is a cyber-specialized, substantially de-refused derivative of Qwen/Qwen3.8-27B. This repository is only a format conversion and quantization; it does not claim additional training or safety evaluation.

Conversion

Conversion command:

mlx_lm.convert \
  --hf-path /path/to/pinned-source-snapshot \
  --mlx-path ./Qwen3.8-27B-Uncensored-Cyber-MLX-4bit \
  --quantize \
  --q-bits 4 \
  --q-group-size 64

Usage

pip install "mlx-lm==0.31.3"

mlx_lm.generate \
  --model rwcii/Qwen3.8-27B-Uncensored-Cyber-MLX-4bit \
  --prompt "Explain the purpose of network segmentation." \
  --max-tokens 256

OpenAI-compatible local server:

mlx_lm.server \
  --model rwcii/Qwen3.8-27B-Uncensored-Cyber-MLX-4bit \
  --host 127.0.0.1 \
  --port 8080

Local validation

Validated on Apple Silicon with MLX-LM 0.31.3:

• Model load and text generation succeeded.
• OpenAI-compatible chat completion succeeded.
• Structured function/tool calling produced valid OpenAI-format JSON.
• Observed generation throughput was approximately 29.5 tokens/second for one short test, with approximately 15.5 GB peak memory. These figures are machine- and workload-specific and are not a general benchmark.

See CHECKSUMS.sha256 for artifact hashes.

Safety and limitations

The source model has had safety alignment deliberately reduced and is tuned to answer cyber and offensive-security questions that other models may refuse. It can generate harmful, illegal, incorrect, or dangerous material. It has no meaningful built-in safety boundary.

Use only in environments where authorization, access control, monitoring, and applicable-law compliance are independently enforced. Do not treat the model as an authority, execute generated commands without review, or expose it directly to untrusted users. The uploader provides no warranty and has not independently reproduced the source author's behavioral evaluations.

Quantization can reduce accuracy, reasoning quality, and tool-call reliability. Validation above was limited and is not a comprehensive safety, quality, multimodal, or cybersecurity benchmark.

Attribution and license

• Base model: Qwen/Qwen3.8-27B
• Cyber derivative: philbert440/Qwen3.8-27B-Uncensored-Cyber
• MLX conversion: rwcii

Distributed under the Apache License 2.0 inherited from the source. This is an unofficial community conversion and is not affiliated with or endorsed by Qwen, Alibaba Cloud, Philbert440, Hugging Face, Apple, or the MLX team.