One-Week Review
#2
by
Reithan
- opened
Recently upgraded to a better video card, so was looking for a bigger model to play with right around when this dropped.
After messing with it for about a week now, I'd like to weigh in.
Pros:
- Very willing to reason for a decent amount of time! Many R1 distills only do 'token' reasoning, just a sentence or two. This model will give you several solid paragraphs of good reasoning! You sometimes need to give it a
<think>prefill to jump-start it, but then, off it goes! - Good use of longer context! Was testing mostly with 24k context, and this model's ability to grab relevant tidbits from the middle of a 24k context was surprising! It did it often, without specific prompting, and incorporated those elements well!
- Very low sycophancy! The 'reduced positivity' it was advertised with is no joke!
Cons:
- SLOPpy: This model CANNOT be dissuaded from digging crescents with its nails, blowing its pupils wide open, making EVERY fabric into denim, and ALWAYS wearing boots. Even banning those words and phrases isn't enough, it'll reword or create slop-variants.
- Falls easily into a rut: This model falls into repetitive patterns of action and wording even 3-4 messages into almost every roleplay or conversation. It doesn't matter if I'm just talking through a concept, doing a roleplay, or writing a story, I have to constantly go back and edit out repeats in its responses, and if I miss even 1, it's not stuck on that thing FOREVER.
- Mean: Not only is this model NOT sycophantic, it takes everything as a challenge, and if a character's personality mentions even ONE, SLIGHTLY, negative trait, the model will amp that to 11...THOUSAND, and that is now the character's entire personality.
I can indeed confirm the model's reaction to flaws. One of my characters has poor eyesight and wears glasses, so the model does everything it can to mention this about 100 times (Okay, I may be exaggerating a lot, but this happens very often) in a single generation...