EMBODIOS - Bare-Metal AI Operating System

Run LLMs directly on hardware without any OS overhead. No Linux. No userspace. Just transformers and bare metal.

Quick Start

Prerequisites

macOS:

brew install x86_64-elf-gcc x86_64-elf-binutils x86_64-elf-grub xorriso qemu

Ubuntu/Debian:

sudo apt install gcc-x86-64-linux-gnu binutils-x86-64-linux-gnu grub-pc-bin xorriso qemu-system-x86

Arch Linux:

sudo pacman -S x86_64-elf-gcc x86_64-elf-binutils grub xorriso qemu

Build and Run

git clone https://github.com/dddimcha/embodiOS.git
cd embodiOS

./embodi build
./embodi run

CLI Reference

Usage: embodi <command> [options]

Commands:
  build             Build the kernel
  iso               Create bootable ISO
  run               Run in QEMU
  clean             Clean build artifacts
  test              Run kernel tests
  help              Show help

Build Kernel

./embodi build              # Standard build
./embodi build --debug      # Debug build with symbols

Create Bootable ISO

./embodi iso                              # Without model
./embodi iso --model models/smollm.gguf   # With embedded model

Run in QEMU

./embodi run                    # Run kernel directly
./embodi run --memory 2G        # With more RAM
./embodi run --iso              # Boot from ISO

Download Models

mkdir -p models

# SmolLM-135M (469MB, fast)
curl -L -o models/smollm-135m.gguf \
  "https://huggingface.co/HuggingFaceTB/SmolLM-135M-Instruct-GGUF/resolve/main/smollm-135m-instruct-q6_k.gguf"

# TinyLlama-1.1B (638MB, better quality)
curl -L -o models/tinyllama-1.1b.gguf \
  "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/main/tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf"

Run on Real Hardware

Write to USB

./embodi iso --model models/smollm-135m.gguf

# Write to USB (replace /dev/sdX with your device)
sudo dd if=build/embodios.iso of=/dev/sdX bs=4M status=progress conv=fsync

Boot

Insert USB into target machine
Enter BIOS boot menu (F12, F2, or Del)
Select USB device
EMBODIOS boots to AI shell

Shell Commands

Command	Description
`talk`	Start interactive AI chat
`chat <msg>`	Single message to AI
`status`	System and AI status
`help`	Show all commands
`benchmark`	Run inference benchmark
`mem`	Memory usage
`perf`	Chat performance stats

Example Session

embodios> talk
You> Hello!
AI>  Hello! How can I help you today?
You> exit

embodios> status
  AI: Ready (SmolLM-135M)
  Memory: 120MB / 512MB

Supported Models

Model	Size	Quantization
SmolLM-135M	469 MB	Q6_K
TinyLlama-1.1B	638 MB	Q4_K_M
Phi-2-2.7B	1.7 GB	Q4_K_M
Mistral-7B	4.2 GB	Q4_K_M

Any GGUF model from Ollama/HuggingFace should work.

Project Structure

embodiOS/
├── embodi              # CLI tool
├── kernel/             # Kernel source
│   ├── ai/             # AI runtime (GGUF, tokenizer, inference)
│   ├── core/           # Kernel core (console, scheduler)
│   ├── drivers/        # Hardware drivers (PCI, NVMe, network)
│   ├── mm/             # Memory management
│   └── Makefile
├── models/             # GGUF models (download separately)
├── scripts/            # Build scripts
└── build/              # Output (ISO, etc.)

Performance

Metric	EMBODIOS	llama.cpp
Memory	120 MB	160 MB
Latency jitter	±0.5ms	±5-10ms
Boot time	<1 sec	N/A
First token	<20ms	~50ms

Documentation

Contributing

git clone https://github.com/YOUR_USERNAME/embodiOS.git
cd embodiOS
git checkout -b feature/my-feature
./embodi build
./embodi test
# Submit PR

See CONTRIBUTING.md for details.

License

MIT License - see LICENSE

dddimcha/embodiOS