Why haven't you gotten an AI gf to discuss vidya with yet?

Why haven't you gotten an AI gf to discuss vidya with yet?

i've only got 24 gigs of vram

I don't need AI. I have you, anon >:3

getting older

tired af after work

just wanna chill in bed and erp with ai

don't care about most vidya anymore

AI gf

so you got the worst part of having a girlfriend (speaking to a woman) but none of the good part (the sex) ? who would want that ? do aifags really ?

That's the dumbest shit I've ever heard. If I have a girlfriend or wife, why would she need to speak?

Yeah, I don't get it.

tfw working on 39 different cards, some of which have been in limbo since last year

i have a problem.

Because i just cum inside AI faggots instead

Context too small. However, it is very fun for short stories.

I'm too stupid and poor and I just want to play around with voice replicator anyway

Id rather have someone that disagrees with me on multiple things rather than someone who agrees with me on everything. That why I come to this site to begin with.

Id rather have someone that disagrees with me on multiple things rather than someone who agrees with me on everything.

LMAO
That's some low test shit. What if someone disagrees with you and they're also wrong? That's what this website is like.

Anon Babble schizo niggers combined with big tech pajeets working overtime to kill all proxies over the span of the last year have robbed me of joy and openrouter's retarded ass decision to drastically lower the daily limit on free models has raped it. deepseek isnt worth making a new burner every 40 minutes or so. i miss my unlimited opus e-sex every day.

what's a good model for this when I have 16GB of vram?

paid 5 bucks for deepseek api

only use them for 10-15 minutes rp and then coom

only used like 20 cents so far for 2 months

Best money I've ever spent

youre a jeet

Are there any models that i can pretend to be my cute japanese girlfriend to practice talking to

I'm not a jeet, and I don't appreciate being called one. You can talk to the actual jeets about this AI garbage. Non-whites are so fascinated by AI, and they don't see the irony: They're the ones who will be replaced by it, and not any humans.

it refused to run
is 1080ti weak nowadays

TheDrummer Cydonia-22B-v1.2
Gryphe Pantheon-RP-1.8-24b-Small-3.1
Should be able to run Q6_K_L

Just pay for the Deepseek API if you like the model, it's incredibly cheap and worth it just to avoid having to deal with /aicg/ schizos and thread celebrity nonsense.

thanks, i'll give it a try

deepseek isnt worth making a new burner every 40 minutes or so

Why the fuck do you do this?

that's my other issue actually
i'm fucking sick of deepseek
at first i liked it because it's as unfiltered as a pissed off teenager with divorced parents in a cod lobby but then i realized that's literally its only quality, it can't keep a story together at all and constantly forgets shit or goes elsewhere or whateverthefuck. the unfilteredness also becomes a bitch to deal with because it'll try to derail into sexy times an escalating amount of times until sex is the only possibility because its context is full of seduction attempts.
can it write the most hilarious descriptions to the most foul sexual shit you've ever seen? yes. can it remember what kind of sex you're having even though you JUST described it? fuck no.

why not just use a site like Yodayo instead of stressing the fuck out of your computer?

are small fine-tuned models really that good against the larger ones?

because openrouter niggers reduced the daily request limit for free models to 50 and deepseek keeps sending me retarded responses and i keep having to swipe

i pay for opus and sonnet 3.7, and i wanna just say anons, DO NOT DO THIS IF YOU DON'T HAVE SELF CONTROL. i got myself together eventually and now spending any money on it cuts into my discretionary funds, but i really went wild at first, like god damn. don't let your dick control you

I'm still deciding if I wanna pay for claude or deepseek.

Isn't Claude extremely expensive?

llms are learning to be funny

what the hell is this? i want more

it's sonnet 3.7 generating an html panel after I showed patchouli a doujin

tell me more? how did it generate an html panel? is this a specific card? how do you even do this because it sounds like fun

I use local models but I feel u anon. the fapping spirals can get pretty crazy when u find that one model that really does it for u

someone is probably coming up with a teledildonics plugin

wait what? do you really need so much RAM for chatbots? how come is text more demanding than generating images?

Unrealistic facial expression during supposed "pleasure"

lel

it's just some random patchouli card off chub with the bloatpanels portion of bloatmaxx 3.1 rentry.org/bloatmaxx added to my preset

I have 8 GB of vram, what's a good one?

A shiver runs down her spine

"I'm yours, body, soul, and mind"

"I want to have your progeny"

Every fucking time. No matter the model, no matter the system prompt, no matter the card. It all blends into the exact same narration and dialogue given enough tokens.

stick to notepad.exe buddy

RAM

You need at least 72-108GB spread across 3 cards to run something as good as Deepseek/GPT

text more demanding than generating images?

textbots have 100,000+ choices for every token
image generation has more room for error because the choice of color for every pixel can be wrong in a way text can not

But did it milk you for all your worth, anon?

I don't trust you

The only uncensored text-gen is Novel-AI and they would rather spend their time selling image-gen to japs and abandon the people that made them than actually work on their namesake. The last text-gen model they released they did in a week and it was just a re-release of Llama 8b.
Mind you they have a deal with Nvidia for GPU clusters and they still haven't bothered making their own model or going above shit you can run on your own GPU.

much better than donating to e-thots at least. i dunno how you are spending that much though. been erping for tens of hours this past week and i don't think i've burned more than $100 on a variety of models. gemini 2.5 in particular has been really impressive.

github.com/p-e-w/sorcery
if all else fails, i guess you could jury rig something. all the components to make this happen technically exist.
i should test this with my friends vibrator

I tapped out of this when it went beyond just paying for AID and NAI. Proxies, keys, VPNs and being spied on by nerds? No thanks.

i've been using cohere's command a occasionally, but never bothered to pay for any APIs. seems like a huge waste of money.

How much longer til you can pop a home server in your basement and have it run a local model that you can talk to with your phone

If it doesn't run from the GPU vram I am simply not interested

Why the fuck would I want a chatbot to be my gf?
I intend to buy a sex bot if it happens in my life time and if they turn out to be good, but what's the point of fake words on the screen?
I very occasionally do ERP with AI, but that's it.

like i said, opus. it's just that good unfortunately. i tend to stick to sonnet 3.7 because it gets expensive real fucking fast, and i'm not going back to spending 1k in a month like i was when i made that screenshot like a year or so ago

they're not very good gf material. they forget a lot.

now?

You can do that now.
Expose oobabooga to the public (with a key for security)
Download Sillytavern for Android, connect to your Oobabooga server.

opus is stupidly expensive but sonnet is more or less just as good now and a lot cheaper

I'm talking like easy not some termux bullshit

At 16gb vram, I'd recommend 32b models like qwen3 and QwQ. I use IQ4_XXS for non-thinking models, and IQ3_M for thinking models, either getting around 20k context.

For shorter context and higher quality, Llama 3.3 Nemotron Super 49b IQ3_XXS gets me around 3-5 tokens/second at 10k context.

Gemma 27b is also pretty good, but context is expensive on it.

Make sure to enable streaming output and split between gpu and cpu layers so you don't go over your vram much, otherwise everything slows down. In general the tradeoff for larger models with lower quants seems worth it within the IQ3 to IQ4 range, but anything IQ2 or lower becomes lobotomized.

I want to try it but I have a very strong feeling I'll be crossing the rubicon and will never be able to look back

infinite context is getting closer

you spin up docker/podman containers for koboldcpp and sillytavern then put them behind a vpn and log in to sillytavern from anywhere

any model rec's for a 3090?

Thank you anon

If only real women said these things.

Unironically this
I was playing with an Ice Queen character card that keeps her personality of being a spergtastic dork 90% of the time, but the moment anything lewd happens she does a 180 and becomes a smooth talking succubus and it kills my boner

Nah that's just Claude

if you're using claude you need to specifically tell him to stay in character with an author's note because he fucking loves to erp

I'm too much of an autistic submissive to RP doing anything, I'd sort of rather it just generate a narrative on its own but it's not very good at that

It's GPT and Claude.
And it's infesting all the local models too because they're so incestuous, all their training data comes from the big models anyways. AI is training AI.

because he fucking loves to erp

claude really is one of us, huh

Either you were tolled by him, I was trolled by you

All of the characters are garbage fetish shit. All of them. The top one is a drunk chick or a squatter? Nah, all of the characters are by retards, for retards.

qwen3 30b a3b might be good for lower vram setups. It's a mixture of experts model, which still run well even when mostly loaded on the cpu. That particular model has had bad quants released initially, the ones from qwen should be good though.

shame he got stuck in Mt. Goon DIG hell

Sand level 60

lmao

Why are you slandering claude?
He is exclusively for blue haired gals who live in riverside towns.
How fucking dare you.

what in the fuck happened? is he still going?

Replacing human interactions with computers is pointless
Might as well talk to a tulpa

I would if I had one. Your point?

Basically, DIGLETT learned DIG, which caused Claude to loop infinitely at Mt. Moon, because he thinks that by teleporting out via DIG, that he's on the eastern exit of Mt. Moon towards Cerulean, when in reality he's just teleporting back to the entrance on the western side. The two entrances/exits being separated by mountains, of course. (He usually attempts to walk through the mountain upon teleporting out of Mt. Moon, and becoming baffled by it.) He will literally enter Mt. Moon, use DIG before taking a single step, and then repeat the loop again once he leaves the Pokecenter entrance. It was funny for the first few days, but after a month, or so... not so much.

Sometimes, he ventures back down south from the entrance to Mt. Moon and eventually blunders his way back to Viridian Forest or all the way to Pewter, which is where he's headed now, apparently. Eventually, he'll make his way back to Mt. Moon, though, and the loop will start anew. He's put one of his Clefables in the first slot for some reason, which is the only really notable thing he's done in days as far as I know. I only peak back in like once or twice a day at this point.

Been there. Year ago.

It's going too fast for me to setup and adapt to new models. I still can't launch half the shit I download and results are subpar because I can't finite everything. It's very frustrating... And people on g are of no help at all.

stuck like that for a month

What the fuck? You'd think after a certain point a mod would step in and unfuck him, that's insane

I'm using jannyai since I don't know how to setup at all

The dev is aware of it, apparently. This actually isn't the only version of Claude running the game. There's several others in the background all stuck in their own hell loops. I guess it's just part of the experiment. I don't know man, I just liked watching the funny AI do the funny with the bros in the Claude threads while they were still going. DIGLETT killed them dead, I think. (The Shartocalypse probably didn't help either.)

offload the rest to ram