Same prompt
Every model is tested with the same prompt. No cherry-picked inputs, no hidden advantages — the prompt is part of the experiment.
fixed input
local_ai // real_hardware // no cloud
AI models, tested on the hardware you actually own.
Local models, real GPUs.
PC Magik Lab is a YouTube series about running AI models on real, local GPUs. One machine, one method, and every prompt and result you can inspect yourself.
Method
How the series measures things
Comparisons are only meaningful when the conditions are fixed. The lab follows one simple rule: keep everything identical, change exactly one thing, and publish everything.
Same rig
All tests run on the same machine — one rig with an NVIDIA RTX 3090. The hardware never moves between tests, so results are comparable.
one machineOne variable per test
Each test changes exactly one variable. When something differs, you always know precisely what it is and why it matters.
A/B, never A+B+CModel reloaded before every run
The model is reloaded before every single run. No warm-state bias, no carry-over from the previous test — every run starts from a clean state.
fresh stateEverything published
Every prompt and every result file is published in the open repository. Reproduce the tests yourself — nothing is taken on faith.
open artifactsHardware
One rig. No cloud.
The entire series runs on a single machine. If it doesn't fit on this rig, that's part of the story.
NVIDIA RTX 3090
The one and only GPU of the series. Every model, every test, every result — on this card.
memory: 24 GB VRAM
LM Studio
The model server that runs the local models and serves them for every test in the series.
role: model server
Local, end to end
No cloud APIs, no rented accelerators. The models run on the same rig you see in every video.
location: the lab, Poland
About
Who runs the lab
PC Magik
Mateusz Piekut — computer service & homelab operator, Poland
computer service
homelab
local AI
RTX 3090
PC Magik Lab is a YouTube series by PC Magik (Mateusz Piekut), a computer service and homelab operator from Poland. It's about a simple idea: AI models don't have to live in someone else's data center. They can run on a real GPU on your desk — and the only way to know what they actually do is to test them.
"Local models, real GPUs."
So that's what the lab does. Each episode runs models locally on the lab's RTX 3090, with the same prompt, the same rig, one variable changed per test, the model reloaded before every run — and every prompt and result file published so you can verify everything for yourself.
See it run on a real GPU
Local models, real GPUs. Watch the series on YouTube, or dig into the prompts and result files in the repository.