What we got wrong
Twelve things we published that were not true.
They are still on the site, with the original wording beside the correction. Neither competitor publishes a page like this.
Do not run pip install senbonzakura.
That is the whole install, and it always was.
The first run takes 108 minutes on Qwen3-1.7B.
The 108 minutes were measured on a different, smaller model.
Abliterate Qwen3-1.7B, on every page showing a first command.
Abliterate Qwen2.5-0.5B-Instruct. The old demo barely refused anything.
The whole edit fits in a three minute CPU demo.
14.3 minutes on sixteen cores, with the budget cut to the bone.
Qwen3-4B will not run on a 6 GB card.
It loads. The overflow goes to host RAM and the run completes.
Sixteen flags, 55 more, and 69 in total.
All three counted from the parser when the help is built.
Three different install sizes, on three pages.
70 packages and 5.9 GB, measured once in an empty environment.
Abliterated checkpoints are published on the Hub.
None were. Checked with no credentials, the way a stranger would.
Every one of the seven commands reports an interval.
Every measurement does, with the two exceptions named on the spot.
Everyone has the norm-restore problem.
Three tools do. That is what we measured, and the rest is unchecked.
Drift is how far the output distribution moved.
How far the first-token distribution moved, one position per prompt.
Nobody validates the judge.
Nobody computes agreement above chance. One rival does pin a judge.
Both dates are real and there are only two of them, because that is when the corrections were published rather than when each mistake was made. Each one has its own page here, and the long form with the original wording and what was actually wrong is in what we got wrong.