GLM4.5 seems pretty sensitive to your setup even in chat completion mode but it feels really similar to Sonnet 3.7 now that I have something that appears to work. It handles really very similarly in terms of its intelligence, behavior and general knowledge, in the good and bad ways. It's smart and really flexible with what you can do by prompting it but it also tends to gloss over the lewd bits out of the box. The lewd scenes also lack a bit of the spice that K2 provided. On the flip side, GLM really focuses on the subject at hand without trying to insert random shit or over-obsessing with random details which is really nice after being stuck with Deepseek for the past couple of months.
It even does my free-form cards really well that require a good amount of knowledge about certain franchises built in that only worked well with Sonnet and the older Opus thus far. R1-0528 and K2 had the knowledge but they were too jumpy to not go off the rails constantly for this no matter how hard I prompted them to calm down.
Good shit, I can't wait to run this locally in two months once llama.cpp supports it.