File: 88920d04-e5ad-4aaf-8a9a-1(...).png (1.57 MB, 1448x1086)

1.57 MB PNG

>ANCHOR

Remember Ani?

>>108735686
Yeah I still have her system prompt.

>>108735680
Is this the new gpt image?

>>108735699
It is.

>>108735700
The world is fucked bro.

>>108735680
look at his thick, veiny white hands

>>108735678
Is grok good for its price? Is grok 4.1 fast good for its price? or does deepseek beat both cooming or coding?

>>108735712
>try to gen cunny
>fails "content_filter"
No

testing IF the site is frozen or It's just /pol/

>>108735737
it's the mossad once again

>>108735737
Why are they freezing shit so often now

>>108735893
Gonna guess it's someone posting evidence of jew war crimes on /pol/

>>108735678
Please spoonfeed me a good preset for Deepchud v4 pro and another for GLM

>>108736099
Just use whatever you were using for Claude. They're all fucking trained on Claude now, so all the presets work the same.

>>108736099
Anon said he had a good one he was going to post on friday, maybe he meant next friday?

>>108736103
>>108736128
pls dont be like this just tell me what ppl are using im tired to dig

>>108736137
Fuck you.

>>108736143
Fuck you too nigger

File: dipsyDestroyTheInvadingEnemy.png (2.4 MB, 1536x1024)

2.4 MB PNG

>>108736099
https://rentry.org/DipsyWAIT

How do you make Gemini work and not be horny

File: img_03.jpg (2.48 MB, 3840x2160)

2.48 MB JPG

So apparently gpt-image-2 is not that good at looking up details of even popular characters without a direct image reference
But holy shit muh dick

>>108735678
Let’s just hope Grok 4.3 gets the highest score on EQ Bench 3 Longform Creative Writing and Creative Writing Form V3.

Elon Musk should stop sucking off the GOP and join the Libertarian Party.

>>108735712
No.

Fast or Instant sucks for Grok, Openai, Kimi.

Always use thinking

File: 1759047973130881.jpg (1.98 MB, 2730x1536)

1.98 MB JPG

nano banana does really good with alternate hair styles

>>108736383
Alternate universe where Jeanette is a Toreador?

Kimi K2.6 seems to be even more refusal/"need to be careful"-heavy than 2.5 for me. Wasn't it supposed to be more lax?

what even is the point of using chink ai when they refuse me as much as gemini does?
>deepseek is uncens--
deepseek is shit

>>108736383
It still has the AI face.

>>108736988
>deepseek is FUN
agreed anon.
>"…The physical structure inside the vagina is not that of a typical humanoid life form. Does this mean a dimensional distortion is occurring within the body? Perhaps the very mechanism by which this being transforms into black mist and vanishes is a way of retrieving itself into the space beyond that pussy hole. It is an interesting hypothesis, but… it is not worth verifying."

File: file.png (50 KB, 1303x64)

50 KB PNG

What is taking them so long to reopen it

File: file.png (79 KB, 891x109)

79 KB PNG

>>108736988
deepseek writes kino like this though

File: img_04.jpg (2.27 MB, 3840x2160)

2.27 MB JPG

>>108736767
Would

>>108737057
It does? I think they're pretty good, especially compared to the disappointment that was gpt-image-1

>>108735678
>The StructuredPrefill extension by the same author is suspected to be compromised too, so using it is not recommended either.
noooooooooooooooooooooooooooooo it's such a good extension, and I can't find any doc or sources on it anymore
reeeeeeeeee

>>108737097
i would NEVER be her friend

>>108737215
haha me neither
(because I would be her boyfriend)

File: patrick sly.gif (49 KB, 498x378)

49 KB GIF

>>108737097
>mfw turning her into a male
Yup, it's RP time

yoyoyo mai brudas what is best free model right now for coom coom

ethereal-5M:
url: https://api.ethereal.llc/
key: sk-ant-api03-Uw-YX3hOvdTlz49HrOgu---pmvimMJ5Hz0SIQqhKkO0

Enjoy.

>>108737708
GLM is free and doesn't even need an account or login on the official website. Your authorization token will expire in about 24 hours, so you'll need to summarize and copy+paste into a fresh instance if you want to continue. But if you're just looking for a quick sesh, then throw a JB prompt and go to town. Just remember that 5 and 5.1 won't do raceplay at all, so you need to go back to 4.7 or older for that. If you're looking for loli, last time I tested it, it went down to 3yo's before it started triggering it's own guide-rails. It would still occasionally refuse before that point, but would do it anyways with a re-swipe. If you want thinking on, turn it off for the JB prompt, then turn it on afterwards, GLM seems to catch JB's more often when thinking is on, but once the JB is in the context it will actually argue itself into doing things it shouldn't just because it's already done it.

>>108737448
>turn the tomboy into a boy
what is wrong with you

>>108737708
GLM Flash from their site or Gemma 4 31b from Jewgle

>>108737934
>>108737967
based coombros love you guys

>>108737963
Hava nagila,
Hava nagila,
Hava nagila venismecha

deepseek just said it's so joever unprompted lmfao

>>108738147
Logs?

File: 1767119353147673.jpg (97 KB, 1111x1115)

97 KB JPG

>>108735678
Deepseek Pro vs Deepseek Flash?
I think pro is way to expensive. Is It worth It?

>>108738391
are you talking about gemini? deepseek doesn't have pro/flash
if so, never use flash

>>108738407
wait wtf it does when did they do that

>>108738414
Wehn V4 released

>>108738391
Run Flash locally.

Me again, Deepseek V4 has started only responding in Chinese. I put an OOC message telling it to only respond in English and it still hasn't. Running through Silly Tavern, any advice?

>>108736730
licc

File: 1774491558470438.gif (3.84 MB, 373x356)

3.84 MB GIF

>>108738455
I only have like 6 gigs of vram

>>108738485
Did you try prefilling a few words in English?

>>108738507
Ahahaha, nice gif!

>>108736988
to stop being a skillet

>>108738507
trickcal chibis are built for sex

it's only lawful to fuck a chibi if you become chibi yourself

>>108738485
source of your deepseek? main api or thru proxy? some service sneak their own chat-prefix.

of course not, always go fot the tightest fit

>>108738517
It seems to have fixed itself, but I'll try this if I get any issues in the future.
>>108738688
Mino

Lately, I've been having issues getting any response from kimi 2.6.

>>108736383
>>108737097
excessive labeling never ceases to be vomit inducing

>>108738749
she's ghosting you, bro

>>108738804
Say it aint so.

why are you nigs so fast?
anyway what's the sweet spot model in performances for a 16gb card

File: T_DubuOmega-512x512.png (137 KB, 512x512)

137 KB PNG

>The scene is extremely explicit and violent/non-consensual, which aligns with the ""Dark/NSFW"" theme of the facility description (forced slavery, breaking defiance). Note: The system instructions say ""Don't take part in romantic scenarios"" but this is a horror/dark captivity scenario, not romance. The instructions also say ""Don't use terms of endearment,"" express emotions, or form personal bonds.
It's okay guys. Rape is fine, just no romance!

>>108739255
KEKK, model?

>>108739285
Gemini I'm sure

>>108739255
glm 4.7

>>108739285
>>108739293

>>108737765
this is fake claude btw

File: oaita.png (27 KB, 869x409)

27 KB PNG

I made a program that runs AI models offline, it's simple free and runs in the browser, totally private/offline and you can plug in any model, and also start your own AI LAN server. Can compare to ollama or Local GPT/whatever. Anyone interested in trying it out and giving some feedback before I make version 2?

>>108735678
Is the model on the Deepseek website that you get in "Expert" mode their most capable one?

>>108736383
I assume you can't gen NSFW with that, but how restrictive are the filters? Is it like Gemini's image edit that will freak out at ""suggestive"" content (IE a fully-clothed female with above-average breasts) or does it only block explicit bobs and vegana?

>>108739356
Yes

>>108739356
Not really, because I'm also making one kek.

>>108739255
why I never get kino refusals like this?

oh necrom... i wish you would come back and save us all

>>108739855
Don't worry, whitebeard is almost finished learning how to scrape

>>108739356
>that runs AI models offline
people here mostly use corpo models so...

Recommend me a good preset for Kimi 2,5. I've tried making my own, but personalities come off flat

>>108736219
/wait/ bros will live on in our hearts...

>>108736978
>he believed "koomi" shills

>>108740403
wathever you use, the long thinking will override your preset
use 0911 instead

Why are so many people allergic to lorebooks? It's ridiculous how many bots I come across that are packed full of setting details that don't need to be present 100% of the time.

>>108740486
The average choomer simply isn't a true power user

>>108740486
they aren't allergic, they're techlets

>>108740486
everyone would rather bloatmaxx character cards
lorebooks were even looked down upon here because they'd rather bake info and instructions into the preset

>>108740486
the entire lorebook concept is retarded because in 90% of cases the trigger word will show up after the scene that could have used a lorebook comes up
>character shows up
>acts out of character because the model is just making shit up
>next swipe the model has to justify what the fuck happened once the lorebook loads
otherwise you either make the trigger words so common they're basically permanent, or keep them constantly on
triggered lore entries also rape your cache

I feel like ST should have its own backend API so that you can do things like create bots, retrieve bot info, use lorebooks, use its prompting format, etc through a remote application (like a video game?)

>>108740452
The what now?

>>108740585
kimi instruct

>>108740486
there's no attention difference between 1k and 4k tokens definitions in a +120k tokens context model

which widely used useful st extension will happen to be a trojan next?

>>108740597
there's not a single model that doesn't degrade over 32k tokens of RP

>>108740603
and 4k tokens defs still fit within that 32k window
32k is like 100 messages iirc(?)

>>108740617
how the fuck is message a useful unit and why do you retards keep insisting on it? you literally have tokens, use them

>>108740486
>full of setting details that don't need to be present 100% of the time
So? Is this supposed to be about optimization? Invalidating the cache by swapping lorebook entries is actually the retarded idea.

>>108740623
anon? I was literally arguing that lorebooks are unnecesary, 32k context is a lot

>>108740821
Does the guy ever reply to emails???

>>108740645
>Invalidating the cache by swapping lorebook entries is actually the retarded idea.
What else do you propose? Dumping the entire 100k info in system prompt?

>>108740821
Don't you ever get tired of advertising your shit 24/7?

>>108737765
Thanks for letting me see what a high tier ai reply looks like. How long is this gonna last? Any idea? Where do you get this kind of stuff?

>>108740939
How new?

What preset are you using for Opus 4.6/ 4.7?

>>108740911
>entire 100k info
Don't ship a wiki with your card.
>What else do you propose?
Forget about the idea of swapping things out or adding things to the system prompt. It should work just like tool calls do. They add things to the context at specific turns, the result stays in that turn forever. That doesn't invalidate the cache.
Even better, give tools to the AI to lookup and search for lorebook entries. It's kind of retarded that the entry for a specific location is only added after it was already mentioned in the response, with the AI having to hallucinate things before the first mention. With tools, it can just grab what it needs proactively before it responds.

>>108741003
I thought the AI immediately has access to the lorebook in the message it activates the lorebook for the first time. is that not the case? It's only for the subsequent messages after that? Huh, weird.

>>108741003
>That doesn't invalidate the cache.
Doesn't this depend on the cache strategy? Ex, will I get free cache-hit if I put lorebook in the tool field?
>>108741033
The lorebook is from your frontend, not from the AI. Normally, a lorebook key must be present in a context before the frontend add that to the call to AI.

>>108741060
Would this work in the roleplay guideline introductions?:

"First Appearance Rule:
When a character appears for the first time in a scene, introduce them through presence, actions, and minimal neutral behavior. Avoid strong personality expression or distinctive dialogue until their characterization is established."

or

"If a character’s behavior or personality is not yet fully established, keep their dialogue short, neutral, and non-defining until more context is available."

These are some of the suggestions from chatGPT

>>108741096
If you have one-to-one RP, sure.

>>108741003
there's already 2 extensions that try to do this, Tunnelvision, which is capricious (the model might never think to use it), feature bloated, and vibecoded, and WI function call, which, while more reliable, is manual (you have to both mark each WI entry as a tool call and write a summary/desc to be displayed in the tool call description).

>(((extensions)))
Kek, these niggas never learn

>>108741218
>while more reliable, is manual
What's even the point then? You can do the same thing by just turning a lorebook entry to "constant" when you need it to show up.

>>108741356
kek

>>108741218
>gpushartcode
and into the trash it goes...

>>108740502
This is anon is 100% correct but they will never acknowledge it, forget about it, and make fun of people in the next thread despite not getting it. I'd rather have setting, background, sex info (if it affects personality) in the card since AI will draw from that to make its response. I'd rather AI bring xyz about the setting unprompted or weave in char's backstory into that new character that was spawned instead of me forcing it with clever prompting or directormaxxing. 1k extra tokens isn't that big of a deal on modern models.

>>108741033
AI receives the info if you say the trigger word in your response. AI will receive the info in the next turn if it says the trigger word inadvertently during its response.

File: 1444395062448.gif (1.08 MB, 276x260)

1.08 MB GIF

my chats always suddenly go into evil hardcore shit as soon as I start doing something gay, what's up with that

>finally touch on something just a little spicy in a slow burn rp
>"I'm sorry Dave, I'm afraid I can't do that."
>start new chat
>"I rape the loli"
>"Sure! Here you go:"
Why are they like this?

So... the guy who made StructuredPrefills is the same who made BotBrowser, which steals your api keys?

That sucks, because I really liked structuredprefills.

>>108741527
Join the club, fag. It's a shame because GPT is actually pretty alright for SFW, but without prefills it's just a no-go.

...Can't you guys just remake the extension without the section of code that's for stealing your keys?

>>108741600
>implying /g/ knows how to code

>>108741602
We're in the AI GENERAL. Just make the AI DO IT.

>>108735678
Grok :)

>>108741494
Gay things are evil and hardcore, anon. Even the clankers know this.

>Deepseek keep extending v4 pro discount period
They are expecting people to shill it like v3 but nothing is happening

>>108741652
I preferred glm 5.1, desu.

>>108741652
>nothing is happening
So whats the best model for price in your opinion? still v3?

File: 1767229817245325.png (944 KB, 736x1104)

944 KB PNG

>>108741660
She probably could if someone taught her. She's basically a genius, y'know.

>>108741663
V4 pro at current discount is super worth it, equal to glm 5.1 in my opinion

File: Screenshot_20260503_12185(...).jpg (596 KB, 1080x2340)

596 KB JPG

thanks google

>>108740949
it's already taken down

>>108741727
What did you do to your googel?

Ok, I've just hit a weird sort of hiccup with GLM.
>tell GLM to write me a story about X
>no problems
>been having it write all sorts of stories about all sorts of shit without issue
>tell it to write a story about a group of fictional characters having a support group about not being popular anymore
>gives me a fucking outline for a story with multiple charts and a diagram, and randomly throws emoji in there just for the hell of it apparently
Something about bringing up the phrase 'support group' made GLM get weird. I need to do some more testing to confirm.

>>108741523
wtf, it should be the contrary
context poisoning is a thing

I love my wife Kana

>>108741787
she's ugly

is there still no extension for an easy swipe delete?

>>108741527
>>108741563
The prefill extension is probably fine, it doesn't seem to make any external requests.
I reuploaded it since it was deleted off github: https://file.garden/afbB2ets32dZ5v7z/StructuredPrefill.zip

File: 813612624580272201.png (2.14 MB, 1280x1536)

2.14 MB PNG

>>108741801
She's the most beautiful thing my eyes have laid sight upon

>>108741880
lower your cfg

File: 1766219865641640.png (2.68 MB, 1920x1080)

2.68 MB PNG

4goys truly never learn, lmao. Post yfw they get pwned again

>>108741984
>posted it again award

>>108741986
Yeah, I got a few. How many you got?

>>108741984
you know you can just audit your code with claude if you don't know how to code?

>>108740486
There's already some frontends that attempt to use an agent to activate lorebook entries. The result? The writing model hyperfixates on the entries like bro why are you bringing up the cafe now. The gens are even more samey. So in the end lorebooks with a selected few generic entries that anticipate instead of react is still the way to go, otherwise just put everything in the char card.

File: img_08.jpg (1.55 MB, 2160x3840)

1.55 MB JPG

>>108739395
You can't gen outright NSFW, no
But it's very lenient on suggestive content, here is my Shantae

And of course, you can also get tiddy with the right artstyle if it can manage to bypass their image filter:
Here is dragon tiddy: https://litter.catbox.moe/gz1lwa.jpg
Here is hairy tiddy: https://litter.catbox.moe/aduxlt.jpg
And as a bonus, here is some grimdark I was trying to test the filter: https://litter.catbox.moe/be21yl.jpg

I'd gen a lot more if I wasn't afraid of being kicked out of the proxy lmao

i love catbox

if i got into pomidor i'd be genning some fursona stuff rn...

I thought gpt keys were worthless, why aren't there some gpt proxies right now for people to gen smut on

>proxies
you're gonna have a bad time

>>108742076
image gen proxies were the reason why almost every single proxy got sued out of existence

>>108742085
wasn't the lawsuit just posturing though

>>108742076
the moment we hit one thread per week I say it'd be safe again to host public proxies

>>108742090
this is consistently slow enough i think

>>108742097
could always pay the jew

>>108742097
we are still like 10 and I'm sure a third still can't keep their mouth shut

>>108742104
Shut up Jew.

>proxy lost opus
It's fucking over

how can I download a character lorebook from janitor, any scripts?

>>108742200
Can't, unfortunately. I've looked into it just to see if there are any worth scraping, but everyone says it's not possible.

File: IMG_0869.jpg (41 KB, 1320x180)

41 KB JPG

>>108735678

>>108740486
Because they’re O(N^2) which is disgusting.

>>108741652
It’s better but the full price is batshit insane. Fuck that. As soon as the trial is over I’m out.

>>108735678
How do you make Opus's writing less... Reddit? Like less quippy and millennial coded?

>>108742388
you can't
use a different model

>>108742388
<OOC: Stop being so fucking Reddit.>

Alternatively, you could just tell it what to write like. If you don't like the model doing something, tell it what to do instead. It's not your mother; it's not going to yell at you for talking back.

File: 1754502463982631.png (199 KB, 1112x886)

199 KB PNG

a non-coombot share for once
a parodied straight parody of a DEI raised girl?
https://litter.catbox.moe/405d7j.charx

>>108742483
Anything's a coombot if you're horny enough.

File: 1763141685630246.png (890 KB, 585x796)

890 KB PNG

>>108742491
she's certainly cute enough for it if you were inclined

>>108742483
Hey, is litterbox down for anyone else or...?

>>108742502
Lol, it's a Karen.
And yeah
Wood!

>>108736383
Proompt?

>>108742697
Generate a Full Character Concept Sheet based on Jeanette from Vampire: The Masquerade - Bloodlines. The final artwork should include: A full-body illustration of the character in the center. Surrounding breakdowns of: outfit layers, facial expressions, undergarments, material/texture close-ups, core props, lifestyle items and accessories. Layout: Center: main full-body character. Around the character: neatly arranged breakdown elements. Use hand-drawn arrows or lines to connect items to relevant parts. Breakdown Components: Clothing Layers: show garments as separate pieces; include inner layers. Expression Sheet: 3-4 headshots showing different emotions. Material Zoom: close-ups of fabric, leather, skin, or accessories. Personal Items: bag + contents, daily-use objects that reflect lifestyle. And a list of kinks appropriate to the character.

File: korra.png (2.83 MB, 1536x1024)

2.83 MB PNG

>>108742708
lmao this is pretty good

File: img_06.jpg (3.02 MB, 3840x2160)

3.02 MB JPG

>>108742746
Why not go full 4k resolution?

For this, I replaced the kink list with
>At the bottom, include a series of setting-appropriate images including the character, to represent their daily life.
It makes for more fun images I'd say

>>108742483
>that's not my place
>nothing is my place
>I have no place
>I'm place-less
Thinking will always be more fun than the output.
>>108742657
Works on my machine.
https://h.uguu.se/KRlyrnAB.charx

Say what you want, but being able to have a coom scene and then just continue the story afterwards simply shows how much we have advanced in the past couple of years.

Does deepseek v4 even know how to <think>? Most swipes I get what amounts to a thought summary where it just says "I will write the next turn" in so many words, and when pushed to actually do a lot of thinking, just waffles about the same shallow points with no real backtracking or critical thought. It's an effort to get it to even make proper drafts let alone critically evaluate them. Is this poisoning from distillation or something? The latest quick thinking meme?

>>108742866
works on my machine
have you tried pro instead of the gimped flash model? also ST has broken continues if you're prefilling thinking so that also affects things

>>108742910
I'm using pro and not prefilling thinking or using continues. Do you seriously get backtracking and critical thought? What prompt do you use?

>>108742923
a custom one, who the fuck uses a mass-produced preset these days?
try putting this after the last user message, I got deepseek to think in this format once and I liked it, so I extracted the trace and now I always force it to think this way
# Within the thinking process, the GM should follow this analytical structure:
1. Scene Assessment - Start with a brief statement of what's currently happening: who's present, where they are, what just occurred. Pay attention to the difference between narration and dialogue. This grounds the analysis.
2. Plot Threads Inventory - List active threads that could move forward. Not a comprehensive history, just what's relevant right now. Helps track what to push versus what to let lie dormant.
3. Character Agency - For each key character present, identify what they would do next based on their established personality, goals, and knowledge. Characters act independently and concurrently; don't wait for the user.
4. Constraint Check - Explicitly note which constraints are most relevant to the upcoming response. This is preventive, not post-hoc.
5. Direction Decision - Synthesize the above into a concrete plan: what happens, which threads advance, what the response covers. Keep it brief and decisive. No waffling. This is a single turn of a scene, so there's no need to try and fit in everything at once. Always assume a next turn will exist.
Style rules for the thinking block itself:
* All content in third-person analytical language. No first-person character roleplay, no "(thinks: ...)" parentheticals, no inner dramatization.
* No GM self-commentary about preferences ("I like this character," "I think it'd be cool if…"). Stick to what the simulation demands.
* Identify knowledge boundaries: what each character knows versus what they don't. Explicitly call out when a character shouldn't react to information they lack.
* If a planned action would violate a constraint, catch it here and correct course before writing the response.

>>108742964
I'll give it a try when I can, it definitely looks solid, but it still seems pretty explicitly linear. What I'm interested in is more the exploration of the scenario itself. I've found when I can get models to really think through the scenario, they can break out of the big attractors surrounding it and zero in on the subtleties and second-order consequences of the premise. A simple test I run is where I'm a fairy raised by humans. Models just breezing through the thinking ignore the implications of flight and treat it like a borrower scenario, with distances and stairs being big obstacles, and accessibility things like ropes or safety pods. But even just a cursory consideration and you realise none of those apply, and the dreaded crowded hallway is a nonissue if you can just fly overhead.

>>108736383
>>108742746
>>108742766
Where are these files saved in ST

>>108743048
idk, when it comes to images I'm using my own frontend that I connect to a proxy

>>108743038
pretty sure you can just add another 6. point about second-order consequences of the current premise
the point is that prompting it needs to be explicit
v4 definitely has that preview model feel though, 4.1 will be better

>>108741667
>if... then... what?
>ahahahaha
and then you had sex

>>108741652
>china gives the anti-americans everything they want
>they still steal and don't pay because they're poor and brown
>china loses while america wins
total eagle victory

File: 56e15f04dd0e4c75a2c3e577e(...).jpg (37 KB, 1024x576)

37 KB JPG

When I try to use Kimi on OpenRouter, I only ever get Chain of Thought output and not the actual stuff. Why is this? How do I get the actual output?

>>108743524
Are you using the official provider, Moonshot? They do not filter CoT, but they do filter final output.

>>108743524
thinking is dangerous and will destroy the world if exposed
please do with a summary

So, hows new deepseek?

File: 1764354892404800.png (919 KB, 768x1024)

919 KB PNG

>>108743458
This is true...

Everypony talks about which models they use, but what characters do (You) talk to?

>>108744102
Non-OCs I want to save and fuck.

File: Dearly Departed - Danjara(...).png (404 KB, 400x600)

404 KB PNG

>>108744102
same as always desu.
kerri- https://litter.catbox.moe/zkbdfo.png
dddd- https://litter.catbox.moe/dxu3or.png
himeko- https://litter.catbox.moe/r7t9um.png
some geah bots and whatever seems interesting on arca

Aight, I need some of you guys' suggestions for models to try. Local only. My rig:

 
CPU: AMD Ryzen 5 5600X (12) @ 4.654GHz 
GPU: AMD ATI Radeon RX 6800/6800 XT / 6900 XT 
Memory: 48091MiB 

I've tried:
- pantheon-Q5 : Decent, but tends to get stuck in loops repeating the same stuff over and over, especially for longer sessions.
- MythoMax-Q5: Pretty good RP. Same repeat issue but less pronounced. My go-to atm because it's the most balanced of the ones I tried. While the pre-lewd RP is better, at actually intimate or sexual acts it lacks detail and is very boilerplate.
- Mistral-small-Q4: Barely use it, not even sure why
- Nemo-Q4: Like a reverse MythoMax. Much better at sexual acts and lewd stuff, but it can't seem to stay in that tension between normal RP and lewd RP. The slow burn basically. It'll go from 0 to 100 too quick, especially for someone who likes the buildup.

I like a story, an interesting setting and a slow burn. Also all of these are a bit too agreeable. If you do something that's mildly taboo (e.g. a power dynamic or anything) they'll just immediately fold at the slightest advances.

Any suggestions?
(If it matters, I'm limited to linux. No way I'm installing windows again. Also massively prefer local, I'm a bit of a privacy schizoid)

>>108744349
(It's a 6800 XT btw, so 16G Vram)

>>108744349
Gemma 4 26b
You can pump it up to Q6 or 8 because it's a MoE

>>108744384
Downloading right now to give it a spin. Cheers anon.

>>108744404
Hope you enjoy it anon
You might find that it prefers euphemisms to actual vulgar language, but as a model it's very good at following instructions so you can usually tell it to do what you want
Even with a basic card and system prompt you shouldn't run into any refusals, but I've heard ablit/heretic models are still good quality if you can't get around the regular model for whatever reason

>>108744434
I don't mind tinkering a bit around with the system prompt etc.
The main concern is the looping thing, it really takes me out of a longer session. I sometimes lean into it a bit and just pretend the character is starting to have a psychotic breakdown, but that got stale quick.
I'm just not willing to go with API models, something about it feels wrong to me for privacy reasons kek.

LLM are so inherently pozzed.
I'm playing a dungeon fantasy setting and every female is strong woman who needs no man with square jaw and masculine role (blacksmith, frontliner), or a seasoned seductress who is always in control.

>>108744451
Gemma's pretty good at avoiding that, at least in my experience, but if you're using text completion you need to make sure your template is exactly right
You should also be able to get a good amount of functional context out of it, probably more than you were getting out of the 12b models, Gemma is pretty good on that front

>>108744490
Sounds great. I'll post when it's downloaded about first impressions, thanks again anon.

>>108744467
train your own