Same prompt
Every model gets the exact same prompt. No cherry-picked questions, no tailored phrasing — one prompt, shared with you, used across the board.
Local models, real GPUs.
PC Magik Lab runs large language models on a physical, everyday GPU rig — no clouds, no render farms — and measures exactly what happens. One rig, one methodology, every prompt and result published for you to check.
The whole point is comparability. Every test in the series follows the same strict rules, so the only thing that ever changes is the thing we are testing.
Every model gets the exact same prompt. No cherry-picked questions, no tailored phrasing — one prompt, shared with you, used across the board.
All runs happen on the very same machine — one workstation, one NVIDIA RTX 3090 24 GB. The hardware never changes between tests.
Each test changes exactly one variable. That is what makes a result readable: you always know what moved and what stayed put.
The model is fully reloaded before each single run. No warm caches, no leftover state — every run starts from a cold, clean load.
Every prompt and every result file is published to the open repository. Read the raw output yourself — no screenshots of what we like best.
Because the prompt, the rig and the method are all out in the open, anyone with a compatible GPU can re-run a test and compare notes.
There is exactly one test bench for the entire series. Same GPU, same server, same machine — from the first episode to the last.
NVIDIA RTX 3090 · 24 GB VRAM
Mateusz Piekut
Computer service & homelab · Poland
PC Magik is a hands-on operator who spends his days in computer service and his evenings in a self-run homelab — somewhere in Poland where the machines never really sleep.
PC Magik Lab is the place where that daily life meets large language models: real local GPUs, real local inference, and measurements you can reproduce with the exact prompts and result files from the repository.
New tests land on the YouTube channel, and every prompt and result file is kept in the open repository.