Same prompt
Every model in a test answers the exact same prompt. If the input changes, the comparison is over.
A YouTube series · running locally
Local models, real GPUs.
PC Magik Lab is a series about running AI models on real, local GPUs — no cloud, no hand-waving. Same prompt, same rig, one variable changed at a time, and every prompt and result file published so you can see exactly what happened.
# one variable per test · model reloaded before every run
$ run --prompt ./prompt.txt --server lmstudio
✓ model reloaded
✓ prompt loaded
→ writing result file… done
$
The method
How the series measures things
No vibes-based benchmarks. Every test follows the same rules, so comparisons across models are actually fair.
Same rig
One machine, one GPU: the same hardware is behind every single run in the series.
One variable changed per test
Each test changes exactly one thing — so you can attribute any difference in the output to that one change.
Model reloaded before every run
The model is reloaded before every run, so no run inherits state from the one before it.
step 01Load the promptidentical across the test
step 02Reload the modelfresh state, every run
step 03Run on the same rigRTX 3090 · LM Studio
step 04Publish the filesprompt + result, on GitHub
The hardware
One rig. That's it.
Keeping the hardware constant is what makes the results comparable — every model in the series runs on the same box.
The rig
Single machine, single GPU — the same setup behind every test.
GPU
NVIDIA RTX 309024 GB VRAM
Model server
LM Studioserves the models locally
Rig count
Oneone variable changed per test
served locally via LM Studio
About
Who runs the lab
PC Magik Lab is the work of one operator, one GPU, and a habit of publishing everything.
PC Magik
Mateusz Piekut
Computer service and homelab operator from Poland. The person behind the rig, the prompts, and the published results.
PC Magik Lab is a YouTube series about running AI models on real, local GPUs. The premise is deliberately boring: one machine, one GPU, and a strict method that keeps the tests honest.
Everything the series produces is public — the exact prompts, and the result files for every run — so anyone can re-run the tests or inspect what actually came out of the model.
See it run on real silicon
Local models, real GPUs. Watch the series on YouTube and dig through the prompts and result files on GitHub.