PC Magik Lab
Local models, real GPUs.
A YouTube series that benchmarks AI models the way a real desktop can run them — on one fixed rig, with honest, reproducible numbers.
How the series measures things
Every test is built to be fair and reproducible. One variable at a time, on the same machine, with everything published.
Same prompt, same rig
Every model is run on the identical hardware and software setup, so differences come from the model — not the machine.
One variable per test
A single setting changes between runs. Everything else stays fixed, so each result points clearly at what caused it.
Model reloaded before every run
The model is loaded fresh for each test, removing carry-over effects and keeping the numbers honest.
Every prompt published
The exact prompts used are shared in the repository, so anyone can reproduce the same inputs.
Every result published
Raw output and result files are made available — no cherry-picked numbers, nothing hidden behind a video.
Real local models
This is about running AI on real, local GPUs — not cloud benchmarks. The numbers reflect what you could get at home.
One rig. Nothing cloud.
The whole series runs on a single, fixed setup so every test is directly comparable.
GPU
A full 24 GB of VRAM, the workhorse for running large local models at useful speeds.
Model server
LM Studio serves the models, giving a consistent local inference stack across every test.
What changes between tests
Tap a card to see what stays fixed and what gets changed. Only one thing moves at a time.
The model
Swap the model and measure how it behaves on the same prompt.
Temperature
Change the sampling temperature and watch how output varies.
Speed & output
Measure tokens per second and compare real-world responsiveness.
A series for people who run their own models
PC Magik Lab is a YouTube series about running AI models on real, local GPUs. No cloud, no abstractions — just a desktop rig and honest measurements.
The focus is reproducibility: the same prompt, the same machine, and one variable changed per test. Every prompt and every result file is published so anyone can verify the numbers themselves.
If you care about what your own hardware can actually do, this is the channel for you.
Local models, real GPUs.— PC Magik Lab
Watch and dig into the data
Follow the series on YouTube and explore the published prompts and result files.