Running AI models on real, local GPUs.
Local models, real GPUs.
A series that benchmarks large language models the honest way — on one machine, with one prompt, one variable at a time. No cloud, no hand-waving.
One variable. Every time.
The whole series is built on a repeatable, transparent method so every test can actually be compared.
Same prompt, same rig
Every run happens on one machine and is driven by the identical prompt, so results come from the hardware, not from noise.
One variable per test
Each test changes exactly one thing. That's what makes it possible to say which setting actually moved the numbers.
Model reloaded before every run
The model is reloaded fresh each time, so nothing leaks between runs and no warm cache hides the real behaviour.
Every prompt published
The full set of prompts lives in the repository, so anyone can reproduce exactly what was asked.
Every result file published
The raw result files are shared too, so the numbers behind each test can be checked and re-run.
Facts only
The series reports what was actually measured on the rig — no invented results, no fabricated statistics.
A single machine, nothing hidden.
There's one setup behind every test. No fleet of cloud instances — just one local rig doing the work.
NVIDIA RTX 3090
24 GB · the only GPU in the rig
All models run on this one graphics card. Its 24 GB of VRAM is what decides what fits locally and how fast it runs.
- GPUNVIDIA RTX 3090
- Video memory24 GB
- Model serverLM Studio
- SetupSingle local rig
One rig
A single desktop runs every test, so comparisons are apples to apples.
LM Studio server
LM Studio serves the models, giving a consistent local inference stack across all runs.
Local, not cloud
Everything runs on-device. No remote API calls — the numbers reflect this machine.
Benchmarking AI the way it should be done.
PC Magik Lab is a YouTube series about running AI models on real, local GPUs. The focus is on honest, repeatable measurement — the same prompt, the same rig, one variable changed per test.
The goal isn't to chase the biggest numbers. It's to show what these models actually do when they run on a single consumer card in a home setup, with everything reloaded and every result published for others to check.
Watch it and dig into the data.
Head to the YouTube channel for the videos, or open the repository for every prompt and result file.