Great model and platform to build on, awesome size compromise without creeping into 70B-class
Compared to ReadyArt/L3.3-The-Omega-Directive-70B-Unslop-v2.1 (3.5bpw):
(5bpw tested): Good logic and creative (but sometimes slightly worse prose) in responses once the (finicky—getting it coherent but not stubborn between swipes) settings are dialed in. Very (if not a bit too) verbose. Great model size, best alternative of the bunch I have tested.
Notes on the others:
Anubis-70B-v1.1 - similar prose, creative output but less logically capable/stable even after playing around with params but good memory retention
AnubisLemonade-70B-v1.1 - much worse logic, not much else to say
Strawberrylemonade-L3-70B-v1.1 - sometimes even better prose, good writing with creative responses. Less logically capable and less stable. Good memory
Cydonia-24B-v4.2.0 - Smart enough in understanding complex concepts, but characters feel less lifelike. Has a tendency to talk/act on behalf of {{user}}
MS3.2-PaintedFantasy-Visage-v3-34B - Decent prose, surprisingly good logic at some situations but a bit unstable, stubbern between swipes
70B models tested on 3.5bpw EXL3. 20B and 30B models tested with 6bpw.