it seems to be the best grok version on the UGI leaderboard, but it's a smart AI which means it's more unwilling (willingness score) to do certain stuff, AKA flat out refusal.
https://huggingface.co/spaces/DontPlanToEnd/UGI-Leaderboard
I personally prefer sillytavern because I like having short sessions where I try to go and make the situation as stupid and unhinged as possible, and since I use free model's on OpenRouter (100 free messages per day~) the AI is pretty stupid it tends to get lazy / go schizo after like 200 messages.
I personally use Mistral-small 3.2 instruct + a futanari fuckventures card I that I like to modify but the problem is that I don't even know what settings I use (I messed with the params and prompt/instruct templates and I add a message to stop adding new characters sometimes) so it might suck / not work with the default settings of sillytavern.
Generally I like making the AI do multiple attempts and if the AI is really easy to guide if you just change your words to guide the AI (you could even make it very explicit like "*I do this and it's going to make her do XXX*"), but if the AI still avoids what you want to do, you could even edit the AI's message to use your words (guiding it) and also continue writing it's reply with your words.
Custom models can be more specific for like horror and stuff, and I think it can still be really good locally if you have 16gb of vram with a 12B q8 model, but I do notice 70b models are smarter (non free model like Sao10K) and it does help with ERP sometimes.
Mistral Small 3.2 instruct isn't a uncensored model but it just works. Nemo Mistral instruct is more uncensored I think but it's dumber, since it's 12B.
I might try "Venice Uncensored" next.
>They are recording your history / honeypot / selling your data
My ISP knows more about my history than whatever french AI company. I'm more worried about being fined for streaming pirated movies / shows off websites like swatchseries.
