Hardware

What she runs on today

No data centre. Luna is built to run two ways: everything on one desktop, or with the model on a dedicated AI server and Luna on whatever machine you're at. Today she's the all-in-one, on a desktop with a 16 GB graphics card, talking to a home server for the smart home and tested on a Mac. The server is the next step, and it's below.

Luna's desk

runs Luna

CPU: AMD Ryzen 9 5900X
Board: Gigabyte B450
GPU: AMD Radeon RX 6950 XT, 16 GB
RAM: 32 GB
OS: Fedora 44, Hyprland (the MewNix Candy rice)
Terminal: kitty, JetBrainsMono Nerd Font

The stack on it

Brain: Qwen3.5 9B (Q4_K_S) on llama.cpp, Vulkan. 16k context, every layer on the GPU, four parallel slots sharing one KV cache
Eyes: The model's vision adapter (mmproj), so screenshots and the webcam work
Recall: Qwen3-Embedding-0.6B for recall by meaning, Qwen3-Reranker-0.6B to sharpen the pick. About 0.6 GB and under 1 GB of VRAM
Ears: faster-whisper (small) with Silero VAD
Voice: Kokoro, af_bella, through kokoro-reader on :8899
Face: Her VRM in an app-style Chromium window, three-vrm, served by her own live monitor

All of it fits on the one card with room to spare. LM Studio works in place of llama.cpp; she picks whichever is answering at startup.

UnraidTower

home server

CPU: Intel Core i5-12600K
RAM: 32 GB DDR5
Storage: Two NVMe drives, small NAS cube case
Network: Dual Intel I226-V 2.5 GbE
Power: APC Back-UPS 1000
OS: Unraid, Docker containers and a VM

What it does for her

Home Assistant: In a container, with its built-in MCP server. This is Luna's smart home: lights, plugs and sensors, local only
Plugs: Two TP-Link Tapo P125M smart plugs, the first devices she switches
Checks: What "how's the tower doing?" asks over SSH with your own keys, and what a scheduled check reads from a text file every few hours

The NUC

serves this site

Machine: Intel Core i7 NUC
Role: Apache web server and reverse proxy for the imbored.me sites
This site: Flight PHP with FlightJS, no database, no analytics, no third-party scripts. The repo stats on the front page come from a one-hour file cache

The Mac

macOS testing

Machine: Apple-silicon iMac, 16 GB unified memory
Brain: llama.cpp built for Metal, with a Mac-sized settings file: one slot, smaller batches, the embedding and reranker servers off so the model leaves room for macOS
Voice: Kokoro on the CPU, played through afplay, with the 44.1 kHz fallback Bluetooth headphones need
Status: Luna runs there, voice included, since October 2026. It's where the macOS side gets tested before it's merged

The laptop

spare

Machine: Acer Predator
GPU: NVIDIA RTX 2080
RAM: 16 GB
Role: The NVIDIA machine, for trying anything CUDA-only

Next: a dedicated AI server

planned

Plan: A headless Debian box running llama.cpp with its own 16 GB card, or an RTX 3060-class card on the Unraid box through an OCuLink dock
Target: A 32B model, which needs 24 GB+ of VRAM or two cards. Luna works on 16 GB today; this is what gives her a bigger brain and keeps the desktop free
Why: This is the shape she's meant to run in: the brain on its own box, always on, the desktop free. Then full home integration, testing the practical side, and later a fine-tune of her own

What you need

Any GPU, or a Mac, or patience

Piece	Needs
The model	A GPU with enough memory for the model you pick: 16 GB runs a 9B with the side servers, as above; a 16 GB Mac runs the 9B with them off; 24 GB+ or two cards opens up a 32B. NVIDIA, AMD (Vulkan or ROCm) and Apple silicon (Metal) all work. Any model with a tool template: Qwen, Llama 3.1+, Mistral, Hermes
Speech recognition	faster-whisper on the CPU, with PyTorch's small CPU build for voice detection. No CUDA download
Her voice	kokoro-reader, with PyTorch for your GPU or the CPU. Starts in seconds since one loaded model serves every app
Short on memory?	Turn the reranker off (under 1 GB back), the embedding server (about 0.7 GB), or the vision adapter (0.6 to 1 GB), or halve the context. A Q4 quant of the chat model saves the most of all
Software	Python 3.10 to 3.13, PortAudio, espeak-ng, git. The installer handles all of it through your own package manager

Set her up