Every test is repeatable. Every result is published.
The exact same input is used across all runs for fair comparison — no variation in what's being tested.
A single factor is changed per test, so you can isolate exactly what affects performance.
The model is reloaded into memory before every single benchmark — no stale state, no cache bias.
Every prompt and every result file is published — complete transparency for the community.
No cloud instances. No abstractions. Just bare metal.
The only rig used for all benchmarks. No cloud, no abstraction — local inference running on bare metal.
A YouTube series dedicated to running large language models on local, physical GPUs. No cloud services, no API calls — just raw hardware pushing tokens one inference at a time. The goal: honest benchmarks, transparent methodology, and open results for the community.
Watch episodes, browse prompts and results.