Running AI models on real, local GPUs — no clouds, no hype.
Local models, real GPUs.
PC Magik Lab is a hands-on series about what modern language models can actually do on hardware sitting in your own room — measured with a strict, repeatable method.
How the series measures things
Comparisons are only meaningful when everything else is held constant. That is the whole point of the series — a controlled way to compare local models on real hardware.
Same prompt
Every test uses the same prompt, so the model — not the question — is what is being compared.
Same rig
One machine, one GPU. The hardware never changes between tests — only the model does.
One variable per test
Exactly one variable is changed at a time, so any difference can be attributed to it.
Model reloaded before every run
The model is reloaded before every run, so no run inherits state from the previous one.
Everything published
Every prompt and every result file is published, so anyone can inspect and reproduce the tests.
One machine. No exceptions.
Every model in the series runs on the same local setup, so results are always comparable with each other.
The only GPU in the series is an NVIDIA RTX 3090 with 24 GB. It sits in one dedicated rig, and every model being evaluated runs on it — locally, on real silicon.
gpu: nvidia rtx 3090 · vram: 24 gbLM Studio acts as the model server for the rig. It loads each model and serves the prompts, with the model reloaded before every run to keep the conditions identical.
server: lm studio (local)Who is behind the lab
PC Magik Lab is a one-person operation: real hardware, a strict method, and everything published for you to check.
PC Magik — Mateusz Piekut
PC Magik (Mateusz Piekut) is a computer service and homelab operator from Poland. The series comes straight out of that everyday work: real machines, real GPUs, and local models that have to earn their place in the rack. No cloud, no marketing numbers — just what the hardware in the room actually does.
Watch it. Check it. Reproduce it.
The videos are on YouTube. The prompts and result files behind every test are in the public repository.