Skip to main content
About

Built in a real homelab.

Hands-on guides on local LLMs, GPU hardware, self-hosting, and the homelab kit I actually run — tested on real gear, written up honestly. No AI filler, no listicles for hardware nobody owns.

Site at a glance

12+
GPUs & Mac Minis tested
40+
Local LLM benchmarks published
30TB
Of self-hosted photos & media
2018
Running homelab since
Our Mission

Why this site exists

The answers I wish I had when planning my own homelab.

The problem

Most homelab and local-AI content online is vague. "Can this GPU run that model?" answers are guesses, outdated, or copied from someone else's outdated post. AI-generated listicles flood every search result. Real, tested data is hard to find.

How we fix it

Every benchmark on this site comes from hardware I personally own and run. Models are actually loaded, VRAM is measured, results are published with the exact configuration used. If I haven't tested it, I say so explicitly.

Coverage

What you'll find here

Four content areas — all hands-on, all running on hardware in my rack.

Local LLM hardware testing

Real VRAM measurements running Llama 3, Qwen 2.5, DeepSeek, Gemma, and Mistral on actual GPUs (RTX 3060, 3090, 4060 Ti, 4090) and the Mac Mini M4. nvidia-smi screenshots, exact quant settings, tokens/sec from my own runs.

Homelab guides

Proxmox clusters, Mac Mini servers, NAS configs, UniFi networking, Pi 5 builds, and the unsexy infrastructure (UPS, cooling, cable management) that keeps everything running 24/7.

Self-hosted services

Immich photo migration, Wazuh SIEM, Nextcloud, Vaultwarden — the practical security and backup patterns I use to actually trust my own infrastructure for family photos and passwords.

Free tools I built

Including the VRAM calculator that estimates exactly how much VRAM you need per model + quantization combo. Free, no signup, runs in your browser. More tools shipping as I need them myself.

Our Approach

How I test and review

One rule: I own it or run it. Everything else flows from that.

Measured, not estimated

For GPU and Mac Mini articles, the model is loaded on the hardware. VRAM and RAM usage are measured (not pulled from someone else's screenshot). Where possible I include token/sec figures from my own runs.

Production, not lab

Self-hosting guides are based on the same stack I run for my own household photos, password vault, and security cameras. If it breaks here, it breaks for my family — so reliability matters more than novelty.

Honest about gaps

When I haven't tested something personally, I say so explicitly and explain what the article is based on (manufacturer specs, community benchmarks, etc.). I never present second-hand data as my own.

The Author

About me

I'm Mustafa — a homelab tinkerer based in the UAE with a background in network and security engineering. I've been running self-hosted infrastructure at home since 2018, and the rack has slowly grown into a serious local-AI testbed: multiple GPUs, an M4 Mac Mini for Apple Silicon ML work, a NAS for ~30TB of photos and media, and a small Proxmox cluster for everything else.

Mustafa.net started because I kept hitting the same wall when researching homelab and local-AI hardware: vague answers, outdated benchmarks, and a lot of "it depends" without the depending-on bit. This site is my attempt to publish the answers I wish I'd had — with specific hardware, specific software versions, and specific results.

Transparency

How this site makes money

Two revenue streams, both transparent. Neither one influences what I recommend.

Display advertising

Google AdSense ads, clearly labelled. Ads do not influence which products I review or what I say about them — they're served on content I would have published either way.

Affiliate links

Primarily Amazon (for hardware I actually own and recommend) and DigitalOcean (the platform this site itself runs on). Affiliate commission costs you nothing extra and never changes what I recommend.

No sponsored posts. No paid placement in rankings. No affiliate money for products I haven't used.

Spotted an error? Want me to test something?

The contact page is the fastest route. I read every message, even when replies take a while. If a guide here saved you from a bad GPU purchase or a wasted weekend, that's the whole point of the site.

Get in touch