Every test is designed to isolate a single variable so results are comparable and meaningful.
Every model is tested with the identical prompt, ensuring a fair and apples-to-apples comparison across all runs.
All tests run on the exact same hardware configuration, eliminating hardware variance as a confounding factor.
Each test changes exactly one variable — typically the model — so any difference in results can be clearly attributed.
The model is reloaded from scratch before every benchmark, ensuring no residual state skews the results.
Every prompt and every result file is published openly, so anyone can verify, replicate, or extend the tests.
All models are served through LM Studio, providing a consistent and reliable inference backend for every test.
Everything runs on a single local machine — no cloud, no abstractions.
PC Magik Lab is a YouTube series exploring the practical side of running large AI models locally. The focus is on real hardware, real benchmarks, and transparent methodology — no cloud APIs, no synthetic benchmarks.
Each episode tests a different model, measuring how it performs on a single, consistent rig. The goal is to help developers and enthusiasts understand what local inference looks like on consumer-grade hardware.
Every prompt used in testing and every result file is published openly in the pcmagik-lab repository.
Watch episodes, explore the data, and follow along.