Qwopus — AI skill for Claude Code

🐙 Qwopus

**Local AI coding agent** — built on Qwen3.5-27B, a Claude 4.6 Opus reasoning knowledge-distillation model.

Installation

# 1. Install Qwopus
pip install git+https://github.com/codespermuted/qwopus.git

# 2. Build llama-cpp-python with CUDA (required for GPU acceleration)
CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --force-reinstall --no-cache-dir

The model (~19.4GB) is downloaded automatically on first run.

System Requirements

• GPU: NVIDIA, 18GB+ total VRAM (e.g. RTX 3090, 4090, dual 5060 Ti, etc.)
• CUDA: 12.0+
• Python: 3.10+
• OS: Linux

GPUs are auto-detected, and you'll see a clear message if VRAM is insufficient.

---

Usage

# Run inside any project folder
cd your-project/
qwopus

# One-shot mode
qwopus "Explain the structure of this project"

# Specify a working directory
qwopus --cwd /path/to/project

Commands

Run `/help` to see the full list.

Command	Description
/help	Show help
/status	Session info
/clear	Reset the conversation
/save	Save the session
/resume	Resume a session
/quit	Exit
!command	Run a shell command directly

Tools

Tools Qwopus uses autonomously:

**Bash** · **FileRead** · **FileWrite** · **FileEdit** · **Glob** · **Grep**

---

How It Works

User input → LLM reasoning → Tool call → Feed result back → Re-reason → Answer

• Local GPU inference via llama.cpp
• Automatic multi-GPU sharding
• Dangerous commands require user confirmation
• Repeated identical tool calls are auto-stopped (hallucination guard)

---

Credits

Model

• Qwen3.5-27B — Qwen Team (Alibaba) · Apache 2.0
• GGUF quantization — mradermacher

Inference Engine

• llama.cpp — Georgi Gerg