Warning

Prototype Stage: This project is currently in a pre-release/prototype stage. The actual coding functionality has not been fully implemented yet as research into optimization is ongoing.

Zweek Code

AI Coding Assistant That Runs on YOUR Machine
No cloud. No telemetry. Just you and your code.

---

What Is This?

A terminal-based AI coding assistant powered by specialized local models. Unlike cloud assistants, everything runs offline on consumer hardware using optimized small models (135M-1.1B parameters).

Philosophy: Use deterministic tools when possible, AI when necessary.

Features

✅ 100% Offline – No internet, no tracking, no data leaves your machine
✅ Fast – Router stays resident (~150MB RAM), responses in seconds
✅ Private – Your code is yours
✅ Lightweight – Runs on older hardware (4GB RAM minimum)
✅ Smart Routing – SmolLM classifies intent (CODE/CHAT/TOOL) instantly

Architecture

Optimized 3-Model System:

1. Router (SmolLM-135M) - Stays loaded, classifies intent with GBNF grammar
2. Code Drafter (StarCoder-Tiny) - Generates code when needed
3. Chat (Qwen3-0.6B-Q8_0) - Answers questions about your code

Key Optimizations:

• Resident models stay in memory (~350MB idle)
• GBNF grammars eliminate hallucination
• Compiler check (cl.exe) validates code instantly (no AI)
• Peak RAM: ~500MB during inference

Quick Start

Requirements

• Windows 10/11, Linux, or macOS
• 4GB RAM minimum (8GB recommended)
• CMake 3.20+
• C++17 compiler

Build

git clone https://github.com/usr-wwelsh/zweek-code.git
cd zweek-code

# Download models (place in models/ directory)
# - smollm-135m-router.gguf
# - Qwen3-0.6B-Q8_0.gguf  
# - starcoder-tiny.gguf

cmake -S . -B build -G Ninja
cmake --build build
.\build\zweek.exe  # Windows
./build/zweek      # Linux/macOS

See QUICKSTART.md for detailed instructions.

Usage

> /help
Show commands and tips

> what is this function doing?
Chat mode answers questions

> add error handling here
Code mode generates fixes

Keyboard:

• m - Switch between Plan and Auto mode
• y/n - Accept/reject changes
• Ctrl+C - Exit

Models

Place these in models/:

Model	Size	Purpose	RAM
smollm-135m-router.gguf	~150MB	Intent classification	Resident
starcoder-tiny.gguf	~200MB	Code generation	On-demand
Qwen3-0.6B-Q8_0.gguf	~700MB	Q&A	On-demand

Download from HuggingFace (GGUF Q8 quantized versions).

Performance

Target: <15 seconds for most operations
Idle RAM: ~350MB (Router + Code Drafter resident)
Peak RAM: ~500MB during chat inference

Status

v1.0.0-alpha - Phase 2 Complete

✅ TUI with FTXUI
✅ Router with GBNF
✅ Chat mode with Qwen3
✅ Compiler-based validation
✅ Command system
✅ Persistent Chat History

Commands

• /help - Show available commands
• /history [n] - Show last n messages
• /sessions - List saved sessions
• /load <index> - Load a previous session
• /clear-history - Clear current session history
• /cd <path> - Change working directory
• /ls [path] - List files in directory (current if no path given)

Keyboard Shortcuts

• m - Switch between Plan and Auto mode
• t - Toggle visibility of "Thinking" sections
• y/n - Accept/reject changes
• Ctrl+C - Exit

Tech Stack

• llama.cpp - GGUF model inference
• FTXUI - Terminal UI
• nlohmann/json - JSON handling

License

MIT - See LICENSE

---

No cloud. No compromises. Your code, your machine.