Very good!
Spent some time testing it on various sfw cards as I usually do. I'm impressed with the output I got so far. Examples below.
1.) My favorite card, works great for testing how creative the model can be dialogue-wise:
2.) A test of model's capabilities to maintain the quirks of speech (reply number 10 was from another L3 finetune):
3.) Another simple test of creativity:
4.) Descriptive text test (impressed with how laconic it turned out to be, usually the models I use are needlessly verbose):
5.) Now for big guns, my most complex, world info-heavy card past 760 messages and with the extended context of 16K. The output is very good. Of note, Celeste seems to lack the tendency of some L3 models to end messages with some 'positive thinking' crap commentary. The output below is continuous, without swipes or regenerating. Never broke the formatting or infobox:
6.) Just for the hell of it, tried to switch perspective to the random city on the other side of the world and see if infobox updates right. It was a success:
Thanks for the detailed review! It's always appreciated.