Recipes · RTX PRO 6000 · Verified on LocalMaxxing
New open models, measured on our own cards, started with one command.
Each recipe pins the image, the weights and every flag, explains why the model runs as fast as it does, and links every number to a Verified run on LocalMaxxing that anyone can check.
House rules
What every recipe here promises.
Every number links to a run
Each speed on a recipe page is a Verified LocalMaxxing run: the canonical prompt, the model's output, the engine's timings and the launch command, public.
Pinned, so it reproduces
Docker images by digest and weights by commit. --print shows the exact command before anything starts.
Gains are confirmed, not hoped for
A tuned setting stays only if interleaved fresh boots against stock show a gain, reported with a 95% confidence interval.
The why comes with the what
Weight bytes per token, bandwidth ceilings, tokens per step: each page explains its numbers, including what did not work.