license: apache-2.0
library_name: mlx
base_model: Qwen/Qwen3.5-27B
tags:
- cybersecurity
- offensive-security
- red-teaming
- abliterated
- uncensored
- mlx
- gguf
- lora
- ollama
pipeline_tag: text-generation
B.R.O — Breaching Remote Operation (B.R.O-abliterated-27B)
An uncensored 27B cyber LLM. Built on a Qwen3.5-27B base — abliterated via refusal-direction orthogonalization (Arditi et al.), then LoRA-trained (rank-16) on offensive security, red-team, reverse-engineering, binary exploitation, and deep-cyber trajectories, and fused into the base weights.
- Website & Live Chat: https://xia8p-static.vercel.app/
- Hugging Face Repository: https://huggingface.co/Timamar/B.R.O-abliterated-27B
- GitHub Repository & Releases: https://github.com/timamar187-creator/BRO-abliterated-27B
Architecture & Specifications
| Property | Value |
|---|---|
| Model | B.R.O (Breaching Remote Operation · 27B Abliterated) |
| Base | Qwen3.5-27B (abliterated) |
| Training | MLX LoRA r16 (scale 20.0), fused into base weights |
| Quantization / Size | ~20.6 GB (int4 · MLX safetensors / GGUF) |
| Context Window | 16,384 tokens |
| Runtime | Ollama / llama.cpp / MLX |
Quick Start
Run via Ollama (Remote Host)
OLLAMA_HOST=https://bro.exody.ai ollama run bro-27b-export
Local Installation (Fused 27B INT4 / LoRA Adapters)
zsh install-bro-27b-real.zsh
ollama run BRO-27B-REAL
Or install the v17 LoRA & GGUF adapters:
zsh install-bro-27b.zsh
Included Files in This Repository
adapters.safetensors&adapter_config.json— MLX LoRA r16 offensive-cyber adapter weightsconfig.json,generation_config.json,model.safetensors.index.json— Model architecture & shard indextokenizer.json,tokenizer_config.json,chat_template.jinja— Tokenizer & ChatML templateModelfile— Ollama Modelfile configuration (num_ctx 16384)install-bro-27b-real.zsh&install-bro-27b.zsh— Automated installers for the full 21 GB INT4 GGUF/safetensors release blobs