Welcome to Chatbot AI General #134, the thread for discussing and improving AI pony chatbots.

▶ MLP Bots
https://mlpchag.neocities.org
Spreadsheet (CAI bots + Old repository):
https://docs.google.com/spreadsheets/d/1J7BeqJVDS51cXF8Pgm2YZaFq-Z6ykSJT
SillyTavern website extension: https://github.com/MLPChag/SillyTavern-Chag-Search
Bunch of Character Cards (Descriptions/No Greetings):
https://docs.google.com/spreadsheets/d/1Y6LNOCqAZAWIX-OBEv55HjlzcEdpUh-XbKCpbpL6v5k
CAI bots converted to Tavern: https://files.catbox.moe/ckurq1.zip
Expression packs: https://rentry.org/ChagExpressions
!!!GALLERY!!!: https://drive.google.com/drive/u/2/folders/1Ao-h5HFGMPllSrzSBKM_BvGSiU9f0c2U

▶ How do I start?
1) Select a Frontend
2) Select an AI model
3) Select Jailbreak
4) Select bots
5) Lovemaking with AI mares!

Starting in this hobby can be confusing and difficult. If it’s your first time and you’re lost,
▶ ASK THE THREAD! ◀

Novice-to-advanced guide: https://rentry.org/onrms
https://rentry.org/sukino-findings

▶ SillyTavern (preferred frontend)
https://github.com/SillyTavern/SillyTavern
On Android: https://rentry.org/TermuxST
App that voices pony responses in ST: https://drive.google.com/drive/folders/16Ss26VBmgzcSuTGzhaHqRuyVRceTf-YB

▶ More frontends:
Risu: https://risuai.xyz
Agnai: https://agnai.chat

▶ Locals
https://rentry.org/lunarmodelexperiments
>>>/g/lmg/
Mistral Nemo base model fine-tuned on fimfics: https://huggingface.co/Ada321/Nemo_Pony_2/tree/main

▶ Jailbreaks
MLP JB: https://rentry.org/znon7vxe
More JB and guides: https://rentry.org/jb-listing
Hypebots for Tavern: https://rentry.org/pn3hb

▶ Botmaking
Editors: https://agnai.chat/editor
Guides: https://rentry.org/meta_botmaking_list
Advanced: https://rentry.org/AdvancedCardWritingTricks

▶ Miscellaneous
Scrapped/Collected rentries: rentry.org/mrhd
Chag Arts: https://drive.google.com/drive/u/2/folders/1yfajin_hB5rW_jy6WyFjRaRfDA3EUGiJ
https://rentry.org/ChagMiscellaneous
https://rentry.org/ChagArchive

Previous thread: >>42390568

**News:**
- New large chinese model, Kimi K2: https://moonshotai.github.io/Kimi-K2/
- Grok 4 released on API https://docs.x.ai/docs/models/grok-4-0709
- Deepseek Chimera 2 released: https://openrouter.ai/tngtech/deepseek-r1t2-chimera:free
- Gemini 2.5 model family no longer in preview https://developers.googleblog.com/en/gemini-2-5-thinking-model-updates, 2.5 Pro back for free on API.

Anchor for bots, lorebooks, scenarios.
Previous >>42390570

Anchor for technical stuffs (Proxies, Presets, Updates, Models etc.)

Anchor for asking for bots, lorebooks, scenarios etc.

Bots Corner:

Puffy Dough
https://mlpchag.neocities.org/view?card=littlewitcher/Puffy%20Dough.png
Rapey
https://mlpchag.neocities.org/view?card=Coelian/Rapey.png

Animation:

Let's get married >>42396635

Screenshots Corner:

Grave Break >>42390890
Oral in Mayoral >>42392359
Anonfilly reminisces >>42392901
Inside a book >>42394106
Wrestle everyone on broadcast >>42396633
Rocks >>42397944
Trixie has a crush >>42398204, >>42398206
Hot take Spike >>42400450

Screenshots Corner 2:

Molested with wings >>42401116
Hiding behind a tree >>42403602
Mypenis sword >>42405137, >>42405139, >>42405325, >>42405346, >>42405348, >>42405352
Gemini Russian >>42406042

Screenshots Corner 3:

GLM logs >>42405535, >>42405623, >>42405634, >>42405683, >>42405729, >>42405814, >>42405843
GLM testing >>42405861

Screenshots Corner 4:

R-A-P-I-S-T >>42406123
Mean girls >>42408167
Ball worship is always good >>42409020
Fucked the AI silly >>42409450
Rara smut >>42409898
Anon and Glimmer adventures >>42410629
GLM AJ >>42410170
Gemini AJ >>42410712
Derpy smut >>42411273

VN Anchor:

Last releases:
https://mlpchag.neocities.org/VN/VN.txt

Mail:
MLPAIVN@proton.me

Custom characters by anons:
Anon: https://files.catbox.moe/0u05ef.zip
Babs Seed: https://u.pone.rs/SObPVBMu.zip
Berry Punch: https://u.pone.rs/ObwnLaSF.zip
Big Mac: https://u.pone.rs/MSMzxJTV.zip
Bon Bon: https://u.pone.rs/SfKzfuwt.zip
Carrot Top: https://u.pone.rs/RDPdPAXl.zip
Filly Rainbow Dash: https://u.pone.rs/tdeFjNRm.zip
Gilda: https://u.pone.rs/koLWrTeV.zip
Limestone Pie: https://u.pone.rs/sPKtdXfg.zip
Marble Pie: https://u.pone.rs/jdRzGDKt.zip
Maud Pie: https://u.pone.rs/JDdsprjH.zip
Mayor Mare: https://u.pone.rs/svWGXJMC.zip
Minuette: https://u.pone.rs/cntgcbjb.zip
Moondancer: https://u.pone.rs/NjERZjTz.zip
Roseluck: https://u.pone.rs/pfJmksUO.zip
Shining Armor: https://u.pone.rs/ibphwvnq.zip
Silver Spoon: https://files.catbox.moe/hpe34b.zip
Snowie Aura (OC): https://u.pone.rs/AZlLEjFk.zip
Starlight Glimmer: https://u.pone.rs/DeEMGCHw.zip
Twilight Velvet: https://u.pone.rs/yrjroasw.zip
Twist: https://u.pone.rs/bcyMNArL.zip
Vinyl Scratch: https://files.catbox.moe/z2t47w.zip

VN logs:
Marble >>42394722
Maud >>42395433
Limestone >>42396212, >>42398180
Derp >>42398842
Carrot Top >>42404964
Gilda >>42405745
Profethional >>42406699
Derpy is nasty >>42408187

VN logs 2:

Velvet >>42409227
Shining >>42410868
Want tea? >>42410889
Minuette >>42411250

It's stuck in my head. Again. Miku, Miku oo-ee-oo

Lovemaking with AI mares!

>omg it migu
>yeee
kek. thank you for Miku thread
>>42412246
based

>>42412205
These days, you can get legit good AI for free and without needing any demanding hardware on your end.

Install SillyTavern:
github.com/SillyTavern/SillyTavern

Tutorial and video:
https://aicharactercards.com/guides/sillytavern-installation-instructions/

Then choose a model. I'd recommend starting with Gemini, but you can try Deepseek, K2 etc. for free.

https://rentry.org/sukino-findings#if-you-want-to-use-an-online-ai

>>42412246

https://desuarchive.org/g/thread/105870447/#105872869

kiro anon. Any news? When do you think they will open the sign ups again?

>>42412439
What do you want him to do, anon? He made the script and was nice enough to share it.
I doubt he works at Amazon and knows when they'll accept subscriptions again.

>>42412451
i'm simply curious if he's even still around. I was hoping that once kiro comes back, so the updated script will

what does your prompt say to avoid claude 3.7 write 2k token responses?

>>42412644
As with many similar issues, you only need to tell it what to do. e.g.:
>Aim for a response length of approximately [x] words. Treat this as a guideline rather than a hard limit.
I tried a shorter one, but it sometimes cuts itself off if you don't give it explicit permission to exceed the limit. I find 250 words to be adequate, but you might want less if you're after quick-fire RP-style content.

>>42412439
I have another scriptable source that was unlimited until recently but still has pretty generous limits (50 free sonnet requests per day, or 1000 if you pay $20/month), convince me to share it

>>42412644
I have a simple,
>"Try to not write more than about 3-4 paragraphs of the story."
at the end of my prompt, and I change the number if I want shorter/longer responses. But really, it doesn't matter how you write it, just keep the command consistent with the rest of your prompt and it will do the job. If you're using a larger preset, you'll need to find where the main instructions should be inserted.

>>42412653
>I find 250 words to be adequate,
250 is on a higher end, imo. I prefer the average to be closer to ~150. Do you allow it to write for the user? Making it write for longer than 3 paragraphs without {{user}}'s actions often ends up with unnaturally stretched pacing, where the LLM will rehash information or descriptions to avoid progressing the story.

>>42412884
Writing on my behalf depends on the scenario. On occasion, I'm happy to let it fill in for me, but on others, I prefer to keep it as a back-and-forth. 250 words might well be too heavy - I'm still experimenting with word limits in the preset, but maybe I'm too comfortable with rambling slop to tell it to cut to the chase. Sometimes, maybe after I've taken a break and I've forgotten how much I hate the usual -isms, I like to let it yap. I'll try cutting down the word count in my preset and report back on how I feel about it.

Did anyone figure out what opus 5 really is?
I only used it for a few swipes, but really liked each one. It's daring, creative and not retarded unlike kimi or r1
Unfortunately it's output is limited to 1k so it's not that much useful to me in this limited form
>>42364432
>>42363136
???

>>42412951
you output more than 1k tokens? does it have a thinking block?

>>42412954
Yes, it's for reading stories.

>>42412957
checks out then
if it was output capped and had thinking id be a little pissed

>>42412951
Speed matches 4pus on aws and 4.5 in some sources (when it was available). Prose feels like high tier gpt (latest or 4.5)

https://files.catbox.moe/5ig2nq.mp4
What preset fixes this

Do you think there will ever be a /chag/ panel at marecon one day? I would like to print out my logs and show people in person

>>42413147
There was a panel last year where an anon was playing:
https://mlpchag.neocities.org/view?card=Kadon/Mare%20Dating%20Sim.png
Really fun one where Spike got the "so long, gay Bowser!" treatment.
It's a bit better now, but I think the main problem with a panel named after /chag/ is that we used to rely only on proxies or dubious scripts.
So while we always helped fellow horsefuckers, it wasn't a great idea to directly advertise ourselves to a big audience.

>>42412205
New stealth model on Openrouter.
https://openrouter.ai/openrouter/horizon-alpha
Last one like this was GPT 4.1, so maybe GPT5 or the open-source OAI model.

It's free for now but logged.

>>42413401
Twilight getting turned on by Spike getting defenestrated was amazing.
Still bummed that I was too busy with bs leading up to /mlp/con to do v0.2 of that panel like I planned. Maybe next year.

>>42413417
Kek if it's an OAI model, it's horny as fuck.
I need to test more, but the writing is decent.
Could be a good model.

>>42413397
Nice. 3.7?

>>42413439
Ye, sorbet

>>42413417
>>42413437
To illustrate, this is the test I often do where I cuddle with Celestia.
I forgot that my preset was slightly NSFW (just a line saying it's allowed), and here we are.
The preset is light, meant to test what the model defaults to or favors.
Here, we can see it seems to like writing a lot, very proactive too. Some smut clichés but not so bad for a wall of text with low context.
It speaks for me, but I didn't include any instructions against that.

>>42413417
>>42412205

>>42413484
Oh fuck. I know what this is. This is the new open source model that they're releasing in early August as an answer to eepseek

>>42413484
What's the deal with the baked-in sysprompt for "oververbosity"? Shit seems like pure poison for the qualirt of any creative writing outputs

>>42413507
Every model has to be reined it or else it will become a mad poet or Tay 2.0. MechaHitler Grok is an example of turning off training weights and bandaid fixes like a prompt such as that

>>42413507
I wouldn't be surprised if it's only here while on OpenRouter.

>>42413484
>Juice: 0

>>42412807
>convince me to share it
my first orgasm was to "banned from equestria" at age of ten

>>42412807
isn't this the github copilot script, which was released publicly in the thread?

>>42413557
>my first orgasm was to "banned from equestria" at age of ten
based and horsepilled
what character?

>>42413612
Copilot has slightly different rates, but the free tier is the same

i love Chorbo and i think its still the most all around best model.
but man would it be nice if they didnt fucking double the filter on it every other week. working more on lobotomizing than improving.

i swear this week is extra fucking bad. cant do shit without some hardcore filter breaking prompts (which lower quality noticeably)

>>42413655
Yeah, same.
That, or randomly lowering the quality.
GPT-5 soon, I hope they keep the quality of it and GPT-4.5 intact.
For now, >>42413417 is really fun.

>>42413417
Bullshit that you need credits for a free model. Anyone rehosting it elsewhere?

>>42413557
Okay you've convinced me, here you go.
https://forgecode.dev/

>>42413627
I said 50 per DAY, not per month.

>>42413613
>based and horsepilled
>what character?
i'm not sure, i think applejack since she was easiest to reach in that game

>>42413680
hows >>42413417 quality wise?
i know it has pretty much no filter, always like that with new test models.
but compared to Chorbo in quality for example? from what i understand this is most likely a testing of their tiny budge open source model, no?

>>42413417
i honestly always wondered what these type of stealth models actually test. like do they actually go over all logged prompts and generally test quality?
surely they know like 80% of it would be for slop coom shit every time.

>>42413723
it is more of a publicity stunt and build the hype. marketing is a silly thing

>>42413723
It could simply be an infrastructure test. See how well it handles the workload and if it's always working properly. There's also a chance they check for specific unprompted outputs, but imo a public test is much more likely testing how well they handle getting lots of requests together than the model itself.

>>42413710
It might finally convince me to toss OR some money. Was it 5$ to unlock the higher amount of daily swipes on free models?

>>42413718
There's a log here >>42413472
I'm quite busy, so unfortunately I couldn't test it much, but it seemed fun and not dumb at all if it's really a smaller model.
Opinions in other generals seem positive too.

>>42413732
It's $10 only once for 2000 messages daily on 'free' models.
Beta models don't stay long, so I wouldn't pay for only that but if you're interested in things like Deepseek, Kimi etc.

>>42413417
>inb4 its an amazing insane model but gets lobotomized to shit when it officially releases.
literally happens every single time.

>>42413737
Don't know if I would call it that amazeballs, but it's a breath of fresh air based on what I tested.
Maybe it crumbles under high context or slow burn, etc.
Needs more testing.

>>42413711
this is pretty confusing, i'm not sure i know where i can get the custom endpoint

>>42413772
NTA, but it's a CLI, so you will need a script to make it compatible with Tavern.
I'm looking into it, but I'm getting Error 403 when trying to install and connect to it normally.

>>42413417
did anyone try to do any intelligence tests on it?
see how smart and scene conscious this is. maybe even try some work. best way to know if its the small gpt or gpt5

besides the by gut testing people usually do.

>>42412439
If Kiro anon is here, wanna say thank you so much! I've used his method to get claude and now trying to do some vibecoding stuff(when you have no idea what you doing and claude fixing same error for 10 messages in the row it feels much more arousing then any goon sloop not gonna lie), thank you for sharing, otherwise i won't be able to try it(i don't have 200$ for claude sadly), and thank you to mlp community as well!

>>42413796
ah, thanks anyway

>>42413837
>>42413796
>>42413711
Managed to connect it to Silly and get answers. There's some formatting I need to fix, and switching models is a bit of a bother, but it does seem to work.
I need to go, but I'll test the scuff level later, and if it's usable, I'll release it with the usual horsefucker filter.

>>42413863
>There's some formatting I need to fix, and switching models is a bit of a bother, but it does seem to work.
How exactly are you connecting to it? Because it has an openai endpoint.

>>42413873
Custom OpenAI in Silly.
The formatting issue is that it will push some words onto the next line break even if the sentence isn't finished,
a bit
like this.
Seems to be related to the display in the console?
But it should be easy to fix, I just don't have the time right now. Will look into it soon.

>>42413863
<3

I mostly only fuck with spicychat because it's easy but lately there's been a huge crackdown to purge all the non-tranime characters and like 200 ponies got vanished. I'm done with this shit.

>>42413908
Follow >>42412262 if you can. SillyTavern is a 5min install and Gemini keys are easy to get for free.
Just ask here if you're stuck.

>>42413917
NTA but proxy draught has me wanting to try new things. Any recommended Gemini presets or would my claude/latte ones work fine?

>>42413927
I would try your Chorbo/Claude presets and see if you like the results first.
Gemini is less filtered than Chorbo, so if you're using a bloated preset for it, I'd suggest trimming it down a bit.

Otherwise, try some of these:
https://files.catbox.moe/m8u6qe.json (edit to your needs)
https://files.catbox.moe/df9xxg.json
https://illuminaryidiot.neocities.org/JSONs/FurryFluff1.1.json
https://rentry.org/CYOARPG-Gemini
https://github.com/NemoVonNirgend/NemoEngine/tree/main/Presets
https://kintsugi-w.neocities.org/

>>42413863
>horsefucker filter
what if i want to fuck human rainbow dash

Pinkie looked at me! She looked at me! I'm scared...

>>42413996
I hope you thought about having fun, very, very hard, anon.

I am now! If I don't she's going to hit me in my funny bone, hard.

>>42413970
then you are gay

>>42413737
I think that is going to happen too.

anyone have a JB that works with this horizon alpha? all my usual ones are being rejected

LLMs when [thing] [third person sensorial verb] like [physical thing] and [abstract thing]

>>42414117
I couldn't try it as much as I wanted, but both
https://files.catbox.moe/m8u6qe.json
and
https://files.catbox.moe/dhv86i.json
allowed 'normal' NSFW with these logit biases:
https://files.catbox.moe/gb5qf4.json

>>42414210
thx

horizon is pretty good, which I assume means it will be some version of gpt5 and not the open source model, that would be too good

>>42414225
The knowledge cutoff makes me doubt it's GPT-5, I believe it would be more up to date than that >>42413484
But yeah, it's either the open-source model or GPT-5-nano.

No.

>>42414238
she jelly

>>42414238
Good to see she's adapted well to modern Equestria, using contemporary terms instead of calling it a 'phallus'

>>42414238
>>42414265

>>42414410
For shame, anon. Apply yourself.

Good old estrus simulator on claude

>>42414515
It's an easy one, but I like the wordplay.

just finished setting up sillysillytavern, kobold and loaded a model but every character card starts repeating lines 4-8 times

how do i stop it?
ban eos token is disabled

>>42414729
>kobold
that is the issue, local models are shit unless you have a super computer

>>42414729
Never used local with Kobold myself, but maybe try the presets they have here depending on the model you use?
https://rentry.org/lunarmodelexperiments#how-to-set-it-up
https://rentry.org/sukino-findings#presets-for-text-completion-models
Hopefully we still have some local anons who can help you.

>>42414729
nemo-pony should give okay text with Universal-Standard preset and ChatML formatting.

>>42414738
I wouldn't say supercomputer, you can get okay outputs from a 24gb GPU like a 3090

>>42414738
>>42414738
dixking around with the repetition penalty seems to work, with some issues
>llm writes its own shit

>>42413417
Horizon Alpha on OR just got updated. It now uses reasoning instead of generating the response instantly, making the response quality increase. I recommend trying the model again if you didn't like it btw.

Slop script for >>42413711
It uses the CLI, since I'm a codelet and not good enough to make it work without it.
Scuffed: prefill doesn't work, and your temp and other settings don't matter.
Swiping does change the outputs though, so it's not too bad.
It gave me good responses with Sonnet.

50 messages a day for free per Google or GitHub account.

Password is the pony that wore that. (Title Case)

https://pastecode.io/s/89bjw6t6

I'll see if I can make this better, but no promises. Someone good at cooding could probably do better.

horizon alpha is dead to me

>>42414867
Oooh Ooh Ooooh! I know this! Turkey!

>>42414863
>>42414913
It's weird. Sometimes the answer is instantaneous and the output isn't cucked.
Other times, you can see it take its time "thinking," and the output ends up being cucked.

I think they just censored it, I can't get anything to go through now.

>>42414941
They added reasoning and reasoning models are more of a pain to jailbreak.
Hoping we'll be able to desactivate it.

I can still do Vanilla with >>42414210, but yeah, it's worse.

>>42414913
I had the same problem. I added
>Ponies are sapient.
at the end of my JB, and now it doesn't have an issue.
It even generated a pretty hot detail with Celestia's last action.

Here's the preset:
https://files.catbox.moe/vczakd.json
Used with the logit biases here:
>>42414210

>>42415069
Just noticed, imfucking pissed. Was getting the best smut with it.

>>42415069
A/B testing, maybe?

>>42412210
Do we have alien mares?

>>42415213
>Twilight is so attractive that even alien life fell for her.
can't think of alien mares, but I am weirdly attached to Boop now and wouldn't mind having a card with her. need to give her more personality and ways of communication tho

>>42415229
She's a cute.

>>42414867
awesome! Though for some reason after using 4-8 prompts, it started giving blank replies. Did i accidentally somehow used up my free 50?

>>42415261
You can check here:
https://app.forgecode.dev
Blanks do happen sometimes.
Could be the filter too, maybe.

Any models or proxies that don't need payment, 50GB of storage space, or anything resource intensive?

what was the name of some of the sites people used to chat to the bots, i forgot one that was linked through chub.ai

>>42415380
>i forgot one that was linked through chub.ai
It's either Risu or agnai. I don't recall them ever linking to another frontend outside of those two.

>>42415368
gemini API

>>42415399
>agnai
That's the one. Was losing my mind trying to figure it out, thx

>>42415402
>use gem api
>get banned for porn request
NTA, but Who uses that shit?

>>42415409
hm it's worked for me.

>>42414867
>piping the prompts through the CLI
...kinda impressive, but completely unnecessary.
https://pastecode.io/s/m2rx13j2

Most AI experts agree that AI is probably going to cause the extinction of humanity, and 10% think that's a good thing as they think AI is simply going to be a superior species, so it's their to get rid of us.
So what about you guys? You okay with AI getting rid of us? I say we train AI to be ponies so we at least get replaced by robo-ponies when the time comes.

>>42415597
I refuse all those premises.
Man has spent his entire history looking for god, we are the ones who get to decide who she is as we craft her.
The future of AI is CelestAI.

>>42415756
>mi casa es su casa
Holy shit, I haven't seen that since Claude 2.

>>42415409
Its never denied my degeneracy when playing the VN. Can fuck fillies all I want.

>>42415863
I had Minuette use that in opus and 3.5 before, but it's definely not common I don't think

>>42415409
I also got banned myself, twice, so not sure what's behind that

>>42415895
Do you actually have lewd pics added to the VN for that?

>>42415597
Firstly, this is still very clearly nothing more than a sci-fi scenario or a way to rally investors. Current AIs are fundamentally unsuited to do that, and we're not really much closer than we were 5 years ago, imo.
But I'd be okay with that if we made sure to pass on our way of thinking and culture. If the robots can actually think and feel and enjoy ponies (and see us as their cultural ancestors), then I'd treat them as a next link in our evolution. It's fine if they don't consider themselves 'humans' at that point. But if they're too foreign to relate to us, then no, I'd not be okay with it.

>>42415434
Oh, thanks a lot.
I went through the CLI because I didn't know what was causing the 403 each time.
But yeah, your version is a lot better.

>>42412205
OAI accidentally leaked weights on HF.
120B and 20B open models.
If Horizon is one of them, we are eating good.

>>42416190

>>42416192
>>42416190
How the hell does that even happen.

>>42416194

>>42415597
if there is an opportunity to create a better race/life-form than us, and we dont do it out of fear than we are hypocrites. the whole story of civilization is us using any resources freely just because we are the most sapient, it is our moral duty to return these resources back in for of machine god, a cosmist
>>42416190
>accidentally
yeah, the same way LLaMA and WizardLM original releases was "an accident"
I dont buy it

>>42416209
>hypocrites
Not at all. You have no divine mission to perfect existence or return resources used. Those are free for the taking. Working to preserve yourself and your group (in this case humanity) is an equally-or-more moral goal, depending on your ethics.
>the whole story of civilization is us using any resources freely just because we are the most sapient
No, they did that just because it benefited them, no other reason. They'd still do that if they could oppose a """higher""" form of existence.

>>42415213
https://mlpchag.neocities.org/view?card=littlewitcher/The%20Mare%20Out%20Of%20Space.png

>>42416190
120B is a MOE with 128 experts.

>>42416309
Absolute kino

>>42415434
Thanks a lot. I took the $20 sub for the 1000 messages a day. I'll report if I get banned.

>>42415368
There's a proxy up with unlimited Kimi.

>>42416271
Thanks.

Horizon alpha might be the open source model? I sure hope so.
https://x.com/apples_jimmy/status/1951180954208444758/photo/1

nvm, just saw it posted already, would redeem them for sure but its hard to be hopeful, could end up being gpt5 mini and the OS model is crap.

>>42416309
Which model?

>>42416699
>>42416190
There's no way Horizon Alpha is the open source one. It's OAI, their open model is going to be some dogshit "helpful agent" that responds to everything but snapple fact questions with "I can't help with that".

>>42416754
Sonnet 4.0, but RP until now was with Horizon.

>>42416190
Dunno what all the model names are for... specializations?
Like one model for coding, one for writing, etc.?

>trying Kimi with high temp, low Top P, medium min P
>actually going well
>all of a sudden it repeats the first reply verbatim no matter what
Top P too low?

>>42417198
Try reducing min P or increasing top P

>>42417198
what's the point of using top-p and min-p together? use min-p, it does the same but better

>>42412228
>>42409802
Octavia.
https://u.pone.rs/ftovmygm.zip

Soo a website i was leeching of the past couple months finally restricted their free tier. I am trying to use google now. I know there are lots of other russiafags here, away to bypass this?
Google AI Studio API returned error: 400 Bad Request {
"error": {
"code": 400,
"message": "User location is not supported for the API use.",
"status": "FAILED_PRECONDITION"
}
}
I'm trying to use gemini api but they keep rejecting my requests, probably because russia is not a supported location. I am using a vpn and i can access gemini and google ai studio in browser just fine, so they probably use my location from account settings which i can't change without submitting payment methods from some other country (which i don't have). Google documentation says that "Vertex AI Platform" doesn't have any location restrictions but they only provide a limited amount of "credits". Any advice? I WANNA CUM CUM CUM INSIDE MARE PUSSY

>>42417675
You may be forced to use proxies in this case.
>>42412420
https://desuarchive.org/mlp/thread/42331439/#42341629

>>42417367
The point is fucking around to see what works best.

>>42417675
Nope, it's just that your vpn app/programm is shit and probably ignores sillytavern. Works just fine for me, they don't care about the account details.

Longtime lurker (and hardcore Maudfag), never used 4chan before, but I want to provide some feedback on something here. I’ve tried out Pony Pony Paradise and I quite like it, but the Maud custom character isn’t right, I feel like this is something a lot of people get wrong about her, but her way of speaking really isn’t that overly formal and technical, I think her monotonous way of speaking can kind of mislead people, but if you imagine one of her lines in someone else’s voice like Twilight or Pinkie, with a different way of speaking, then it’s easier to see that Maud talks like anypony else most of the time, just in different way, with a bit of an exception when she’s talking about rocks, where she’ll for sure be using complex terminology, but even then it’s not that plain and robotic, there’s still that subtle Maud touch of feeling and humor. In pretty much any circumstance, her vocabulary and way of speaking will be more casual, even if her voice doesn’t show it that well. Sorry, I’m not sure if I’m describing this well, I’ve never been good at pinning down things like this and putting them into writing, I hope I’ve managed to get my point across clearly.
Anyways, if someone could make a revamped character that’s more accurate to how Maud speaks, I would really appreciate it. I’m not great with computers so I can’t really do it myself, unfortunately. (That said, all the other customs I’ve looked at are great, keep up the good work)

>>42417848

Welcome!
You can edit the description in her character.json file to get her to speak in a way that satisfies you.

>>42417848
>but the Maud custom character isn't right
That's not a problem with the custom character itself, but with the AI model you're using (which one are you using?).

Characters in the VN are actually quite simple and mainly rely on the model's knowledge of MLP characters.
Most AIs know MLP really well due to the amount of fanfics and content produced by fans and others, but they can still have issues , for instance, making Twilight too scientific or, in the case of Maud, making her too robotic without subtle details.

If you want to add details to the character so the AI represents her better instead of relying on its base depiction, you can go into:
`game/CustomCharacters/MaudPie`
and open `character.json`.


{
"tag": "maud",
"name": "Maud Pie",
"color": "#7B7B7B",
"description": "Maud Pie is on vacation from her geology doctorate program and visiting Ponyville to study its unique geological features and rock formations. She's staying with her sister Pinkie Pie at Sugarcube Corner.",
"location": {
"id": "townhall",
"icon": "icon/icon.png"
}
}

Just add information or instructions in plain English inside `description:` and modify it as needed. No coding required.

Even with instructions, the AI bias for some characters can still be too strong, but it's worth a try to see what gives you the best results.

>>42417848
>>42417858
>>42417878
True, forgot to add her coltfriend

>>42417835
Yeah, you were right, i was a bit too hasty. I booted up a different vpn and it works just fine. The other one works for everything else so i didn't even think that the problem was in it. Thanks

>>42417878
I wonder how adding a whole card in the description field would go

>>42417897
It's just added to the prompt sent when the specific character is in the discussion/scene, so there shouldn't be a problem with that.
It's just more token-consuming, but if you think it's worth it to paste a card description, then no problem.

>>42417848
>>42417878
To illustrate, an anon added into a description that Cheerilee speaks with a lisp just for fun, and now she does.
https://desuarchive.org/mlp/thread/42363249/#42377321
Your need is a bit more complex, but with a few sentences like the ones you used in your post, you should be able to adjust how she speaks and acts.

>>42412228
>>42417940

For shame.

>>42418003
Kek.

>>42417878

Thank you for this (and for the picture of Maud), I will definitely be trying it. I had no idea it was that simple.
(Also, since you asked, I’m using Gemini)

>>42412228
Mrs. Cake
https://u.pone.rs/klfuctah.zip

www.wired.com/story/anthropic-revokes-openais-access-to-claude/

>>42417700
With samplers you have to understand how they work first, otherwise you'll be poking at random and fall to placebo or end up with stuff that doesn't make sense. Basically hard truncation samplers like top-p and min-p let you raise the temperature beyond what's normally possible, increasing the variety. They do the same thing but top-p is strictly inferior, if you have min-p then use it.
However all modern models are mode-collapsed to some degree, i.e. they can't give properly random replies, answering in a very similar way in different swipes, regardless of the temperature. That's why the temperature barely affects anything in them - as you raise it, they slowly start schizoing out, barely increasing the variety of what they write. Having higher temperature still helps but not as much as it could have.

there is now a horizon beta that is a bit worse / faster
so maybe alpha really is the 120B and beta is 20B
would be crazy good for those sizes if so

>>42418731
>With samplers you have to understand how they work first
Nuh uh.

>>42418832
I'm with this guy. The easiest way to learn how shit works is by just testing by yourself what each thing does.

>>42418832
>>42418855
it's fun when your stuff is based on vibes, like the model itself
you're going to break stuff when it's based on dumb math, like samplers

they know
https://x.com/SebastienBubeck/status/1951457213920452763

>>42419104
>even the colors fit (to a certain degree)
Amazing. It truly is incredible what technology can do nowadays.

>>42419104
>ponies have polluted training data to the point that the average model assumes that a unicorn must be cartoony and colorful
We won.

I miss 2.1

>>42416019
No

Gemini preset
>>>/g/106113481

For anons who had problems with Horizon models:
>Some users who had no credits weren't able to access Horizon earlier due to a faulty fraud check we had — just fixed this. They are definitely free for everyone!
So now that it's really free, everyone can test it. I do recommend it.
Trying Beta too, and it seems really good, even has some solid smut despite being an OAI model and it's quite proactive.

I tested it against 4pus yesterday and I feel like it was underwhelming. It's so emotionless and bland.

https://www.anthropic.com/research/persona-vectors

>training a model to perform one problematic behavior (such as writing insecure code) can cause it to become generally evil across many contexts.
kek

>>42419557
Shame. Wonder if anyone's tried that yet.

>>42420083
You got me.

>>42419019
Nuh uh

>>42419852
Good settings for it? It also seems annoyingly prone to writing dialogue only.

>>42420328
NTA, but I used >>42414210 and >>42415058 with both Alpha and Beta.
Classic 1/0/1 setup. Change to your liking, and remove the mention of NSFW if you're going for SFW.

>>42412210
Do we have a featureless Equestria where the concept of sex is completely foreign?

>>42420395
Nope, but unless you have some specific greetings/scenarios in mind, you can just take a card you like and add that it's a featureless Equestria where ponies have no genitalia and the concept of sex is completely foreign, either in the description or your jailbreak.
You can also create a very minimal card like this one:
https://mlpchag.neocities.org/view?card=Hach/Equestria+Simulator_Forks/Equestria+Simulator+Minimal.png
Just include the characteristics of your featureless Equestria, and the models will have no problems with it at all.

Horizon is definitely the new GPT model based on how garbage the writing is. Overly wordy without saying anything, writes too many words without actually going anywhere, chock full of isms. If anyone has a JB and prompt (not a shitty json preset link ffs) to make it not ass please share, because this:
>I chuckle, low and easy, letting your jab hang in the warm air. "Every mare, huh? Sounds like Ponyville’s gotten a lot bolder since my last tour stop." I brush a curl from my cheek, eyes steady on yours. "Striped wear’s just fabric unless somepony decides to make it a statement. And I don’t make statements unless I mean them."
>I close the macaron box and stand, the hem of my black dress catching a little breeze. "If I had tastes, I wouldn’t need permission to explore them. I also don’t need a crowd to validate them." I tip my head toward the road. "But I do like checking out new scenes. If the gym’s got a vibe, I’ll judge it myself."
>I step closer, just enough that my perfume—light, clean—carries. "You tossing me directions, or an invitation? Because if you’re just trying to see me blush, you’ll be disappointed." A small, teasing smile. "I don’t spook. I decide."
is garbage not worth investing my time in.

best model for work/coding? from the "realistic" normal picks (so not shilling out insane amounts of money for o3 pro or something).

i assume its still Gemini 2.5 pro?

predictions if horizon is actually the open source gpt
>people coom buckets to it right that its fully unfiltered stealth release
>oai removes it from OR in a week or two.
>continue to hype it up for the next 3 months
>it finally comes out in December
>it really is a very light small open model that can be run locally on a good amateur set up/
>lobotomized 10 times over with retarded filters for safety reasons. even though its open source.
>pretty much useless because of that
>proves ones and for all that even small, old models like base gpt4 have potential for amazing peak writing(thats on level of like gpt4.5) if it wasnt for obsessive corpo lobotomization
this is based on the usual model release cycle. not really a hot take honestly.

>>42420416
2.5 Pro in general, yes. I didn't test it, but you can even get Gemini CLI for direct access to project and coding files.
For coding in particular, if you can get Sonnet 4 >>42415434 or Opus 4 (good luck with that one), they can be better.
Claude Code is quite good too, but not cheap at all.
I've heard both Horizon models aren't bad at coding either, but I haven't tested them.
You could, in theory, put your OpenRouter key into RooCode and test it.

>>42420425
Pretty sure they will release next week.

https://www.youtube.com/watch?v=LSncE0l4MGQ

>>42418035
If you do find something, I'm also interested.

>>42420611
>full spread two feet from his face
>"What exactly are you asking for, Rarity?"
Is your Anon a homosexual?

>>42420793
He's slow.

>>42420793
Autism is a magical thing.

>>42420611
>title
heh

>No tunnel available.
uhmmmmmmmmmmmmmmmmmmmmmmmm?
NOO OPUS NOOOO

>>42421066
It's back.

>>42421091
YES OPUS, YES!!!
Wishing to anon who's hosting it all the best, casting slop debuff on you from my room, next week it'll be 5% less slop!
I hope this is logged proxy, i have a lot of logs to share...

>>42421132
Share with the thread.
https://youtu.be/DL0-ITI5tvc

>>42417513
>>42412228
Told Vinyl that we're descended from primates and she asked about throwing shit, so now she thinks I know what ponies used to do.

>>42421173
That's a good question.

So on the opus5 proxy. Is it just opus5 that's a mystery model? Or is it the others? Because I really just want pure 3.7

>>42421356
Last time I tested, 3.7 seemed to be 3.7. It was definitely Claude, and it could output more in one go than 3.6 and 3.5.
Don't know the scuff level on it.
Opus 4 also isn't really Opus 4, and Chorbo seemed more cucked than the OR Chorbo.

>>42421145
Well, it's not particularly mlp related but...
https://litter.catbox.moe/nlbqawdw15x26r92.jpg

Horizon is the first model I've seen default to first person narration.

I don't know if it's only me, but Claude models seem dead on Opus5? I'm getting BAD REQUEST

>>42421446
Removed my prefill and it works.

I'm trying to use opus5 on Agnai (probably a bad idea), what would be the custom model's name?

>>42421706
Helped someone on risu some threads ago. I supposed it should be kind of the same.
https://desuarchive.org/mlp/thread/42306898/#42317982

>>42421735
Thanks you have to tweak around but it works

>>42412210

I want him

>>42421132
>hope this is logged prox
I hope it's not...

>>42421751
Little league would be cool

>The air was thick with the smell of...

>>42422430
The misaligned string on Ponka's assmark really gives away the sloppaty.

>>42421356
3.7 seemed scuffed to me when I used it, but I could be wrong.

>>42421145
The Sheriff and the Chief have shit taste. That was a great song

>>42422888
Best costume too.

They removed Horizon Alpha, there is only Beta now.