Local models, real GPUs.
Every test is designed for clarity and reproducibility. Same prompt, same rig, one variable changed at a time.
Identical input prompts are used across all tests, ensuring that differences in results come from the variable being tested — not from the prompt itself.
All tests run on the exact same hardware configuration. No swapping GPUs, no changing system conditions — the only variable is the model.
Each test changes exactly one parameter — model size, quantization, context length — so the impact of every factor is isolated and clear.
The model is fully reloaded from disk before each test to eliminate caching effects and ensure consistent, fair comparisons.
All input prompts and their corresponding result files are published and accessible — complete transparency, full reproducibility.
No cloud instances, no abstractions. Every benchmark runs on physical consumer-grade hardware in a real local setup.
A single, consistent setup — no moving parts, no hidden variables.
PC Magik Lab is a YouTube series dedicated to running AI language models on real, local GPUs — no cloud, no abstractions, no hidden infrastructure.
Every episode tests different models under controlled conditions on the same hardware, publishing all prompts and results for full transparency.
The goal is simple: give the community honest, reproducible data about how local models actually perform on consumer-grade hardware.