More discussion about bitch wrangling Mistral Small 3.2 please, just to cover all bases before it's scrapped.
I've tested temps at 0.15, 0.3, 0.6, and 0.8.
Tested Rep pen at 1 (off) and at 1.03. Rep pen doesn't seem to be much needed just like with Rocinante.
Responses are still shit no matter what, but seems to be more intelligible at lower temperatures, particularly 0.15 and 0.3, however they are still often full of shit that makes you swipe anyway.
I've yet to try without min_p, XTC, and DRY.
Also it seems like it's ideal to limit response tokens with this model, because this thing likes to vary length by a lot, if you let it, it just keeps growing larger and larger.

Banned tokens grew a bit and still not done;
>emdash
[1674,2251,2355,18219,20202,21559,23593,24246,28925,29450,30581,31148,36875,39443,41370,42545,43485,45965,46255,48371,50087,54386,58955,59642,61474,62708,66395,66912,69961,74232,75334,81127,86932,87458,88449,88784,89596,92192,92548,93263,102521,103248,103699,105537,105838,106416,106650,107827,114739,125665,126144,131676,132461,136837,136983,137248,137593,137689,140350]
>double asterisks (bold)
[1438,55387,58987,117565,74562,42605]
>three dashes (---) and non standard quotes (“ ”)
[8129,1482,1414]

Extra stop strings needed;
"[Pause", "[PAUSE", "(Pause", "(PAUSE"
Why the fuck does it like to sometimes end a response with waiting for "Paused while waiting for {{user}}'s response."?
This model is so fucking inconsistent.