Welcome to Chatbot AI General #188, the thread for discussing and improving AI pony chatbots.▶ MLP Botshttps://mlpchag.neocities.orgSpreadsheet (CAI bots + Old repository):https://docs.google.com/spreadsheets/d/1J7BeqJVDS51cXF8Pgm2YZaFq-Z6ykSJTSillyTavern website extension: https://github.com/MLPChag/SillyTavern-Chag-SearchBunch of Character Cards (Descriptions/No Greetings):https://docs.google.com/spreadsheets/d/1Y6LNOCqAZAWIX-OBEv55HjlzcEdpUh-XbKCpbpL6v5kCAI bots converted to Tavern: https://files.catbox.moe/ckurq1.zipbonziPONY: https://github.com/maresmaresmares/bonziPONY!!!GALLERY!!!: https://drive.google.com/drive/u/2/folders/1Ao-h5HFGMPllSrzSBKM_BvGSiU9f0c2U▶ How do I start?1) Select a Frontend2) Select an AI model3) Select Jailbreak4) Select bots5) 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/onrmshttps://rentry.org/sukino-findings▶ SillyTavern (preferred frontend)https://github.com/SillyTavern/SillyTavernOn Android: https://rentry.org/TermuxSTPony TTS app supporting SillyTavern: https://drive.google.com/drive/folders/16Ss26VBmgzcSuTGzhaHqRuyVRceTf-YB▶ More frontends:Risu: https://risuai.xyzAgnai: https://agnai.chat▶ Localshttps://rentry.org/lunarmodelexperiments>>>/g/lmg/Mistral Nemo base model fine-tuned on fimfics: https://huggingface.co/Ada321/Nemo_Pony_2/tree/main▶ JailbreaksMLP JB: https://rentry.org/znon7vxeMore JB and guides: https://rentry.org/jb-listingHypebots for Tavern: https://rentry.org/pn3hb▶ BotmakingEditors: https://agnai.chat/editorGuides: https://rentry.org/meta_botmaking_listAdvanced: https://rentry.org/AdvancedCardWritingTricks▶ MiscellaneousScrapped/Collected rentries: rentry.org/mrhdChag Arts: https://drive.google.com/drive/u/2/folders/1yfajin_hB5rW_jy6WyFjRaRfDA3EUGiJhttps://rentry.org/ChagMiscellaneoushttps://rentry.org/ChagArchivePrevious thread: >>43398466**News:**- New Deepseek V4 Flash, new V4 pro soon™ https://api-docs.deepseek.com/updates/- Kimi K3 weights released https://huggingface.co/moonshotai/Kimi-K3- Opus 5 released https://www.anthropic.com/news/claude-opus-5- Gemini 3.6 Flash released https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-6-Flash-Model-Card.pdf
Anchor for bots, lorebooks, scenarios.Previous >>43398467
Anchor for technical stuffs (Proxies, Presets, Updates, Models etc.)
Anchor for asking for bots, lorebooks, scenarios etc.
Bot corner:>Thunderlanehttps://mlpchag.neocities.org/view?card=Anonymous/Thunderlane.pngScreenshots corner:Love Your Mares >>43398514Applelestia >>43399375Twilight Loves a Monkey >>43399958Bat Infestation >>43400775Armpit-Farts Cunt >>43400984AJ Blows >>43401077Kimi Smut >>43402846Britbong Octavia >>43403716Strange Cutie Mark >>43403874
Screenshots corner 2:Reverse Summoning >>43406118Parasprite Abuse >>43408398Luna Dating >>43408733Yeet >>43409786Break the Windows >>43410219Looking at the Wrong Mare >>43410281Heartbreaker? >>43410489Secret Herd >>43411313Flirting with Lings >>43411829
Screenshots corner 3:Balance! >>43412577I Was Born in Them >>43413149Great and Powerful >>43414887Magic Trix >>43415220Just a Prank, Glimmy >>43416324Royal Scratcher >>43416695, >>43417221, >>43417656, >>43417803
>>43419480What is the best model / setup for lovemaking with AI fillies? Asking for myself.
Screenshots corner 4:Terrorbeast >>43416717Fable NSFW >>43418731Apples on a tree >>43419259
VN Anchor:Godot Versions:>Websitehttps://equestrai-ponyponyparadise.neocities.org/>Windows>Main files + Data folderhttps://drive.google.com/file/d/1OEiMXN7o_QG32shGnqmy0QndVDT6nYbF/view?usp=sharing>Main files onlyhttps://drive.google.com/file/d/1Fhe-OeIlNIYFFGx2FlnpuxcIkYdOFGh2/view?usp=sharing>Linux>Main files + Data folderhttps://drive.google.com/file/d/1Vszzn8TTfXFWSQeKiFdNAAQGXfcReTJ3/view?usp=sharing>Main files onlyhttps://drive.google.com/file/d/1LPlIstcgQjyXqhOl4L_ZqMPeUQW5Xn2-/view?usp=sharing>Android (Beta)>APK + Data folderhttps://drive.google.com/file/d/1WB08tguZa1PD1frJKdo7IxlGsp4XR0Wp/view?usp=sharing>APK onlyhttps://drive.google.com/file/d/1nPOSGufXi_-PjmjaKES9waJ6HZZOcL9V/view?usp=sharing>Mac>Full Gamehttps://u.pone.rs/lcblgjjf.zip>Update Onlyhttps://u.pone.rs/qolpovci.zipBug reports or requests:https://scrumboy.com/je1fqcsjSource code:https://gitlab.com/ponyponyparadise/equestrai>TTS VN mod release v3https://desuarchive.org/mlp/thread/43324592/#43333706https://u.pone.rs/xxkjrsje.zip>Impression Mod V2.1:https://desuarchive.org/mlp/thread/43346142/#43354798Nightmare AU (Renpy/Godot)https://equestrai-ponyponyparadise.neocities.org/mods#nightmareModG5 Worldmaphttps://equestrai-ponyponyparadise.neocities.org/mods#g5ModAnons' mods (Dashie mod, Nightmare AU, new cities, Calendar):https://equestrai-ponyponyparadise.neocities.org/modsCustom characters:https://equestrai-ponyponyparadise.neocities.org/charactersVN logs:Evil Glim Glam >>43398908Rock Hard >>43400034, >>43400102Loyalty Test >>43400050On Train with Lyra >>43400067Estrus >>43400172, >>43400782, >>43401330, >>43403934
>>43419506VN logs 2:Rarity's Little Brother >>43404099, >>43404102, >>43406486, >>43406491Raripiggy >>43405362, >>43405363Queen Celestia >>43406171Celestia's Bullying >>43407454, >>43408181
>>43419510VN logs 3:Anon Stoned Adventures >>43407907, >>43407909, >>43407915, >>43407917, >>43411972Pout >>43410446Fluttershy Not Now >>43412354Barn Doors >>43412531Die by Cute >>43414243
Guide for Newcomers:Install SillyTavern:github.com/SillyTavern/SillyTavernModels and Explanations:rentry.org/sukino-findings#if-you-want-to-use-an-online-aihttps://mlpchag.neocities.org/Ponies.txt
>>43419502SillyTavern (guide here): >>43419521Then for the model, DeepSeek V4 Pro is still cheap and not very strict with filters using basic presets.GLM 5.2 if you need a boost in intelligence. I believe it will allow it too, but I'm less sure since I don't do foal stuff.Other anons can correct me on this.
You know, for someone who—*vomits all over {{user}}*
Warning: Self-Indulgent Fluff.
>>43419525What about running locally? I have a 3090
YOU, YES YOU. TAKE THIS DOG AND PUT HIM IN YOUR GROUP CHATS AND POST THE RESULT.https://chub.ai/characters/Anonymous/barking-dog-595c8fd7a3eb
>>43419571i know this dog irl small world
>>43419548The general is mostly API-oriented because textgen models are quite demanding, and those you can run locally are generally less smart and knowledgeable than big ones.Still, we do have some local anons.From what I know, your best bet with a 3090 would be Gemma 4 26B Q4 or Gemma 4 12B.There's a guide in the OP, and also:https://rentry.org/Sukino-Findings#if-you-want-to-run-an-ai-locally
Okay so Deepseek doesn't know Rainbow Dash's parents at all. Had to prompt it with their names and it still got it wrong.
>>43419637Kek. Poor Foal.>>43419547Cute.
>>43419506another thing I noticed is that in custom start scenarios, the VN will always introduce characters in alphabetical order no matter what was written in the custom start scenario context box. For example: I write that the first character I want to interact with in the scenario is rainbow dash and then select everyone in the mane six to appear in the scenario but no matter what, applejack will always show up first. Is this hard-coded into the VN or is there something in my prompts that I have to adjust?
>>43419636Deepseek's just generally very stupid, it'll mess up the physical descriptions of even the Mane Six.
>>43419656It's because of a shitty line in the custom start prompt that I should have deleted earlier (it was mostly there for testing):>2. Begin with EXACTLY ONE visible character: {{primary_name}} (tag: {{primary_tag}}).Remove that line and it will be good. It will be removed in the next version.
>>43418800just FYI, if you're doing noncon/underage/etc with this preset on "fable", it's not actually fable
>>43419915NTA but if you get refusals it's not Fable actually; Sonnet will give you raw refusals even with a lot of context but Fable will go along with it. It's really evident when you're being offloaded to Sonnet.
>>43419919you are not using fable, fable will always refuse those things. there's no such thing as "offloading" on the API unless you explicitly enable fallback models, you are being served a mixture of chinese models probably.
>>43419915Seconding this. Plus the writing in the screenshot is not how fable writes. Could of course be context poisoning or a card/preset issue, but>Somewhere down the street Is peak chinese 'ism. Plus the real tell for me is that there is zero Not X But Y and all that other dogshit claude loves so much.
>>43419637Which model were you using?
>>43419801that seemed to do the trick, much appreciated, anon
>>43419506New version.>Windows>Main files + Data folderhttps://drive.google.com/file/d/1OEiMXN7o_QG32shGnqmy0QndVDT6nYbF/view?usp=sharing>Main files onlyhttps://drive.google.com/file/d/1Fhe-OeIlNIYFFGx2FlnpuxcIkYdOFGh2/view?usp=sharing>Linux>Main files + Data folderhttps://drive.google.com/file/d/1Vszzn8TTfXFWSQeKiFdNAAQGXfcReTJ3/view?usp=sharing>Main files onlyhttps://drive.google.com/file/d/1LPlIstcgQjyXqhOl4L_ZqMPeUQW5Xn2-/view?usp=sharing>Android (Beta)>APK + Data folderhttps://drive.google.com/file/d/1WB08tguZa1PD1frJKdo7IxlGsp4XR0Wp/view?usp=sharing>APK onlyhttps://drive.google.com/file/d/1nPOSGufXi_-PjmjaKES9waJ6HZZOcL9V/view?usp=sharing>Mac>Full versionhttps://drive.google.com/file/d/1DzKEasuhhatilCPlXJPUJYtJ8e7QG_Oq/view?usp=sharing>New Features>Ability to manually change display order of lorebooks & individual entrieshttps://scrumboy.com/je1fqcsj/t/157Done.https://u.pone.rs/bpzgwucs.png>You can now decide and set the timeout limit yourselfhttps://u.pone.rs/hlhopwou.png>When starting with Custom Game, you can choose to start at Eveninghttps://u.pone.rs/vddurkvj.png>Dice rolls are now visible in the context menuhttps://u.pone.rs/knzcvhaq.png>Android UI updatehttps://scrumboy.com/je1fqcsj/t/156https://scrumboy.com/je1fqcsj/t/148https://scrumboy.com/je1fqcsj/t/147https://scrumboy.com/je1fqcsj/t/146https://scrumboy.com/je1fqcsj/t/145These should be fixed for Android. It's not the version I main, so if any anon uses it a lot, I'm open to feedback.>Bug fixes>Fixed some reroll or undo cases where bugs could occur or where the music didn't reset to the initial location>Loading save deletes/overrides personashttps://scrumboy.com/je1fqcsj/t/154Fixed.>Setting the date to any time before the default date prevents the vacant house from becoming acquirable until the default date is reachedhttps://scrumboy.com/je1fqcsj/t/156Fixed.>Add a confirmation prompt before exiting a conversation with Escapehttps://scrumboy.com/je1fqcsj/t/155Done.>Bad instruction for Custom Start >>43419656Fixed.>Problem with Gemini sometimes giving only two lines of dialogueShould be better now. Let me know.>Dreams are bypassed without giving an error message.Fixed. It should now tell you what went wrong.I'm really drunk right now, so I hope I didn't forget anything or fuck things up.If there's something you really, really want soon, just reply to this.
>>43420329Thanks for your work as always!
>>43420329Great work, anon! Can't get enough of this VN.
>>43420329I forgot, but it should also fix the problem where Dynamic and Dynamic+ didn't activate when visiting locations alone.Don't forget to log bugs and requests at https://scrumboy.com/je1fqcsjThere's less chance my drunk ass forgets about them if you do that.
>>43419506what are some good temp/top p/top k setting for GLM 5.2? I keep getting >x is y of x was doing z reddit tier responsesshits annoying
>>43420329What should I download if I want to update my current godot build on windows? Main files + data folder or main files only? Not really good at these kind of things, sorry
>>43420362i fucked around and found outtemp 0top p 0.9top k 50
>>43420368main files and data folder if you haven't updated for a whilejust overwrite everythingsaves are in a different locarion
>>43420377Got it, thanks man
>>43420378welcome
>>43420145Deepseel
Between Kimi 3 and GLM 5.2, which one do you think is the better model? or do you think one model has a specific thing that does better than the other?
>>43420329>You can now decide and set the timeout limit yourselfAs the Anon who asked about that, it's nice to see it added, though I sadly won't be using it after all.Local models (specifically Gemma 12b and anything above) are just too big to fit on my 12GB GPU with the game simultaneously running, and thus run like utter crap.So even if I increased the timeout, it'd still be a ridiculously long wait for the prompt to process, and it also lags my whole PC in the meanwhile.Hopefully the timeout adjustment proves useful for someone else, because I feel kinda bad for otherwise wasting your time on adding that. T_T
Looking for advice. For doing HiE crap, would you say doing it during season 1/2 or season 6/7 would be better. I feel each have their own pros and cons and I can't decide.
>>43420906I personally lean towards S6 because I wanna be able to preen Twilight's wings and Starlight's a fun character. Ultimately though I don't think there's any big reasons to go in either direction other than earlier lets you partake in moments you enjoy from the show whereas you're probably gonna be more slice-of-life-y if you set it later in the show.
>>43420906Why not both? I have my own autistic timeline that included Anon arriving around the start of s2 and emerging from the Everfree around halfway through the season. That way if I want to RP events like the Canterlot Wedding or the Crystal Empire I can, while most of the setting is assumed to be in the last seasons, of course, this varies between cards.
>>43420906Depends on what I want to roleplay. I've done S1/2, I've done late seasons. I've done 1000 years in the future and past. Just do whatever serves your intention best.
>>43420906Neither. Headcanon the stuff you like from later seasons and include them into the Season 1/2 canon. Like Cozy Glow but hate the friendship school? Keep her around as a filly attending the normal schoolhouse. Like redeemed Starlight? She's visiting ponyville and is staying in the library (NOT THE CASTLE PLAYSET). Fucking, even Shining Armor and Cadance needs consideration. Technically, Shining Armor is no longer a member of the royal guard, but the prince of the Crystal Empire. However, nearly everyone keeps him as captain of the royal guard while keeping him married to Cadance. Honestly, you just have to pick and choose what you want to be canon.
>>43420906Can you even keep a model in S1/2 without it inevitably pulling knowledge from the later seasons?
>>43420906Just... what u like more? I like Faust's Equestria so I do that. Easy. I don't like the writing tropes that would prolly emerge from the later season shit anyway, though I guess I can't rlly say much in regards to that cus none of it was for me>>43421076Wasn't a particularly big issue for me. Whatever issues ther eare r all fixed by minor editing of the AI outputs. Which u should be doing anyway wink
>Trying to get Twilight to do lewd things to an asleep human.>Every time I try, she just acts incredibly respectful.I don't know if I should be impressed or annoyed.yes the pp got hard.
>>43420906Why choose? My AI Equestria is all over the place, with Twilight being a unicorn in a library and Discord still being stoned, yet I have the kirin and the Crystal Empire existing wirh Flurry Heart as an adult.Long as you give the AI your own lore it doesn't much matter what's in the training data.
>>43420906Nightmare AU
When you need a shitty setup for your smut.
>>43421690
>>43419485I don't know what's happening, but some providers have insane price drops for GLM 5.2 on OpenRouter right now.-95%https://openrouter.ai/z-ai/glm-5.2#providers
>>43420480K3 is smarter, but I do prefer GLM for the writing and price.I need to test if the providers here are good too or scuffed/quantized somehow >>43421786
>>43419924Wrong. You're just getting injections.
>>43421786I stopped using GLM 5.2. Its endless repetition just pissed me off.
>>43422122Get better.
>>43422142What a helpful comment.
>>43420343I always had the trouble of when visiting locations alone the ai wouldn't get your persona or the character filesevery time I did they were completely absent from the logs at least, have to keep spoonfeeding the ai everything if I wanna go alone
>>43422185Which version (Windows/Linux) are you using? I will look into it more and send you just the fix to see if it's good on your side?
>>43422190I'm on Windows
>>43422146Make your preset smaller, play around with sliders, ask it to write things in a different language first and then translate them, try new cards, etc.
>>43422185>>43422199Can you try with:https://u.pone.rs/hmpjcjty.zipIn the logs, persona should be under:>## Player PersonaAs for character sprites, if you're alone and call Applejack, for instance, the scene where you call her and she appears won't have her list of sprites because we have no way to know that the AI made AJ appear before its output.But the next question/text you send after her appearance should have her full list under:>## Characters Who Joined This Session
What models do you guys like for generating short form (40 total messages or less) porn logs? I'm willing to pay up to $20 a month for AI.
>>43422321Put something like $10 on OpenRouter and try either DeepSeek V4 Pro:https://openrouter.ai/deepseek/deepseek-v4-pro (Make sure to use DeepSeek as a provider in SillyTavern)or:https://openrouter.ai/z-ai/glm-5.2 (There's a massive price drop from some providers, but they still need to be tested >>43421786).They should be good enough for quick smut, and $5-10 a month should be more than enough.If you want smarter models or more, or you have a scene where you want more complexity, you can try switching to something like Gemini 3.1 Pro or Opus 4.6.But they are more expensive.For Gemini and Opus, a preset like Chaggy: https://u.pone.rs/flkjwwqb.jsonShould be good enough for basic NSFW.
>>43422306that seems to have helped, thanks anon
>>43422321https://openrouter.ai/apps/sillytavernCheck prices for each, they differ significantly. In general more expensive means better. Higher number mostly generally better, except where censorship went up like Opus-5, or the odd shit-model like Opus-4-7.If something has a flash/normal/pro version then look at prices to see which is the dumber-but-faster-and-cheaper and which is the "good" version.Look into caching, shit's easy to setup if your preset doesn't prevent it, and cuts costs massively, at least for bigger stories or ones where its a fast back and forth instead off ah ah mistress and then 2000 lines of ai answer.
>>43419506I got a custom start scenario inspired by a fic I read and liked that I use for quick & easy clops or slowburns>Humanity found out that it was not alone years ago. A portal was opened, revealing the existence of Equestria, land of the Ponies, and a much larger world. Soon, a partnership between the two races was formed. All was well. With the staggering ratio of mares to stallions (70 - 30), somepony thought they could make some money to hire out human males to mares on request, providing a wide range of services, from mares who seek simple, relaxation, companionship, romance, intimacy, to horny, kinky mares who really need to get it out of their system, all the while maintaining utmost discretion and anonymity should either be required by the client. It all started from an idea, born from a newspaper article of the Equestrian Times when they interviewed Lyra Heartstrings — a unicorn that became well known for being the first pony to be in a relationship with a human being — as she spilled the beans. The idea was quite simple: if humans were potentially the next best thing since sliced bread, why not share the joy to every mare for a fee? That’s when Gentlemen for Mares was created. "A Gentleman for Mares must be capable of adapting to the needs of any mare, always expanding his repertoire and make each customer a happy one—for a happy customer, is a repeat customer." The player is one such human male and he has been hired out as a Gentleman.
>>43422350I remember reading and liking Gentleman for Mares something like 12-13 years ago, and some spinoffs too.I should give them another try to see if it's actually good or if younger me just had shit taste.
>>43422337I do need it to be smart, so I'll be looking into Opus and Gemini. Thanks for the preset.
>>43422437>Opus4.6>4.8>>4.7(retarded)>>>5(censored)For presets less is usually more, I'd treat his preset as the absolute max size. And get that caching setup if you're going to pay out the ass for top models
>>43422437It's not like DeepSeek and GLM are totally retarded either. I would still give them a quick try if I were you.
>>43419506>>43422350>>43422382I find the story to be quite a fun read, I'm so late to the party because I really only got into mlp fics and rp last weekAnother log
Some GLM presets I found, if someone wants to test:https://rentry.org/evening-truth-glm-52-darkhttps://file.garden/amZZRFkR9Rqi5BhC/Pitclaude.jsonhttps://files.catbox.moe/2ck8ov.json
>>43422507Do you choose the ponies you go with, or do you let the model decide?
>>43420329can you not change your night residence if you pick the quick start option?it's only giving me the library
>>43422869Quick Start is based on a scenario where Twilight is the first to host you, so it's the night residence by default.You can change it, but you first need to visit the location with the pony "owner" in it.For instance, visit Fluttershy at least once at her cottage, Applejack at Sweet Apple Acres, etc.After that, you can go to the menu, click Night Residence:https://u.pone.rs/yrqmtnhb.pngAnd change it:https://u.pone.rs/jdhzltti.png
>>43422454I found 5 to be way too filter-happy as well, but I did test 4.8 and 4.6 and enjoyed my time with them. I didn't like 4.7 at all. Will be trying Gemini next.>>43422456Maybe after I'm done trying Gemini. I did use Deepseek around when 3.2 came out and found it a bit too dumb for my liking. I've never used GLM, though.
>>43422913>I found 5 to be way too filter-happyIt literally inject this on any kind of nsfw. I'm usually really good at getting around filters, but opus-5 just doesn't work.
>>43422928>It literally inject this on any kind of nsfw. Only with Anthropic as a provider. Amazon Bedrock and Google Vertex don't have this.
>>43422934ah yeah sorry, I'm piggybacking off of my claude sub, I always forget about those providers
>>43422847I choose the ponies that appear in the scenario intro but other ponies can appear if it makes sense for them to do so since I have character management set to dynamic+
>>43419506may I ask what are the chances of the VN one day being able to import and use SillyTavern presets? I get that there are special formatting rules for the character sprites, locations, etc but I'm sure those can be edited into SillyTavern presets
I can't decide between nanogpt, openrouter, and smol, please send help!!! I NEED to fuck my mares again!
>>43423237I'm pretty happy with NanoGPT. 60m tokens a week serves me well for deepseek v4 pro thinking and glm latest. Dont think its got latest kimi in the subscription though
>>43423237I'm not sure about nano right now since GLM 5.2 is cheap as fuck on ORThere is also chinkrouter >>43418820
>>43419506Will be in the next update.I added a way to resend the output the AI gave you to have it rewritten.Ask for a rewrite in a certain style, for instance, or to remove some fluff, etc.
Telling ponies about kappas.
>>43419919>>43419924If my first impression of Fable wasn't actually Fable does anyone have a more concrete idea of what it could've been? If it was a Chinese model than I absolutely would start using it because I liked the output it gave of my filthy non-con stories. Trying to wrangle opus 4.7 and 8, something I have easy access too, while possible is tedious as hell. Flurry Heart being absolutely destroyed by a dragon is completely fine but a Hazbin character forced into porno shoot to cringe opus says. Otherwise what Chinese model is someone having success with non-con you recommend keeping in my mind I prefer stories rather than roleplay.
>>43423962Maybe they have kelpies.
>>43423992https://equestripedia.org/wiki/Kelpie
>>43423973Have you tried asking the model? Mine was a Mistral variant.
>>43424089NTA, but asking models isn't that great of a solution because models don't know what they are and many will hallucinate.Still, Claude will always say it's Claude but can say it's a previous version of Claude.Gemini will often say it's a large model from Google.Chinese models can say they're Claude.Old ones used to say they were Gemini.Also, some providers can add instructions asking a model to always identify as another.
Hey is it possible to use the Impressions mod for the VN with the newest update?
>>43424165Good question.It was my goal to analyze it and see how to integrate it into the game, so I can give you a definite answer later today or tomorrow at the latest.If you have time, don't hesitate to test it and report back.
Meta released a new local model:https://huggingface.co/unsloth/Muse-Glimmer-30B-GGUFThey will also open-source their big model, Muse Spark 1.2.
>>43424310>glimmer>Glimmer>GLIMMERNo! They can't! They wouldn't dare!
>>43424322It's a glimmy glam world, anon.
https://ponepaste.org/6642Very, VERY, fun scenario to try out. Getting to experience ponies sperging out while doing sexo is quite the experience
>>43424438There's a video too if you want:https://m.youtube.com/watch?v=y5KjXz37TYcI did the reverse some time ago and had Fluttershy work for a sex line. It was fun and cute.
>>43424475
>>43420329So... How do I install the android version? Running the apk after extracting to the root folder did install the game but it's a gray screen with 0 characters
>>43424650You need to do that:https://desuarchive.org/mlp/thread/43311664/#43322107
>>43424657I see, had to ask Claude The permission wasn't showing up on the normal window, had to look for special access then force feed the appThanks! Now I can finally continue being sweetie belle's little brother ... Barely have time to play.. work work...
>>43424680On this note, where are saves located?Can I just port it over, not a log file?Log files tend to have the history cut off at some points, would rather a save
>>43424700Saves on Windows:>C:\Users\<name>\AppData\Roaming\Godot\app_userdata\PonyPonyParadise\dialogic\saves\Saves on Linux:>/home/<username>/.local/share/godot/app_userdata/PonyPonyParadise/dialogic/saves/Saves on Android:>/storage/emulated/0/PonyPonyParadise/saves/
>>43419506>>43424165First, I'm not the original author. I'm the main VN dev.Here's a version of the Impression mod compatible with the new release >>43420329. I'm calling it 2.2.For now, this is just a quick release to keep Anon's work compatible. I'll still look into the best way to integrate it properly.Windows:https://u.pone.rs/zirzliau.zipLinux:https://u.pone.rs/bbryugli.zipJust unzip the correct version in the game folder and launch the PonyPonyParadise-Impression launcher. It can coexist with the vanilla one, so you can switch back to the classic version if I fucked something up.For the Anon working on this: if your work is based on the previous version, keep going—I'll update it if needed.Here's the original post if you want to see what it does:https://desuarchive.org/mlp/thread/43295554/#43302412
>>43424310>SparkHas anyone tried it?
From all the ponies, glimmy would be the one to summon a random human.
>>43425265>Does toad taste like chicken?Dunno, but frogs taste like chicken.
>>43425280Rule of thumb is:white meat - chickenred meat - beefFish - fish
>>43425305Frogs and alligators are chicken-fish IMO.
slowburnfags : how do you handle longer stories? i grow tired of asking my ai to do a synopsis every 50k tokens or so. i want to try out a really really long story now without worrying about context rot (sillytavern)
Why is Kimi like this?Prefill was just "I need to work on the reply,"
>>43425983I didn't realize Kimi had such strong opinions on the matter.
>>43425938>i grow tired of asking my ai to do a synopsis every 50k tokens or so.That + lorebooks is kinda your only option unfortunately. Even if some models can remember more, the quality is always going to start suffering at high token counts.
>>43424916dude you're the greatest mofucka ever
>>43419547Not the biggest wing-boner fan, but it was cute.
>>43424725Thanks!
>>43425983Prefilling Kimi's thinking is fun, but I couldn't get the schizo or fun tone from the thinking to carry over into the outputs, which often stay banal.If you find something, please share.
>>43419506anon is not going to survive
>>43426381I like them
>>43426381>>43426614Based wing fetishist.
>>43426381Are there seriously Anons that don't take care of their mare's wings or horn if they have them? Seems inconceivable to me at this point not to.
Still pretty sad how we never got a final Silly update just to say that the proxy is going to be kill indefinitely. As it is now, the last thing he said was that he hadn't been feeling the best. He could actually just be dead for all we know.
>>43426841In my defense, she just got them.
>>43419506if I use the sprites of Pony Pony Paradise for a video or a fangame, who should I give credit for in the... well, in the credits?
>>43427434Most of the core sprites were ripped from Boorus and DeviantArt and such. Themselves mostly based on or ripped from the show (with still a lot of work involved; I'm not diminishing that).Since it's supposed to be an /mlp/ project that stays between the 30-40 autists of the general, I admit I didn't bother with credits.There are also custom characters that are made by different anons with different sources.So if you want to use them, you don't have to credit PPP at all. I even prefer if it's not advertised that much.If you do want to credit the original source, I guess you can do a reverse image search on Derpibooru or Yandex to see the original author. But I wouldn't be surprised if the majority of accounts are dead or dormant now.
>>43427478>So if you want to use them, you don't have to credit PPP at all. I even prefer if it's not advertised that much.Eh, then I just credit "/mlp/" in general, if that's fine, if not, I just don't give credit at all (thought I usually prefer to not pretend I made shit I didn't, and these sprites are saving me A LOT of work)
Surely we'll have new Deepseek or Gemini 3.5 pro Soon™
>>43427478>Since it's supposed to be an /mlp/ project that stays between the 30-40 autists of the generalThis shit hit the mainstream fandom Twitter months ago
>>43427752There was one post about it from what I've seen, and it wasn't traced back to the thread because it's not mentioned on the Neocities.
>>43427478>the 30-40 autists of the generalthere's over 100 sillyspoon api keys, and not every anon from the general even got in
Thoughts on big lorebooks? I made some convoluted one after getting inspired, but it's 6000 token without even any openers. It feels nice having it broad and in depth enough so the world works good enough that you're practically unrestricted when it comes to picking a character, but 6000 token is just so much. On the other hand we're no longer in 2024, and the big models have 100k+ usable context now, so does it even matter?
>>43427900Back when I first started using LLMs I thought lorebooks would only activate when key words were triggered and I had no idea I was using 30k tokens a prompt on a bloatbook
>>43427900Are the entries always on? Because if they trigger on keywords only, being 6000 tokens isn't that much of a problem.>the big models have 100k+ usable context now, so does it even matter?Models still start to shit themselves around 30-40K, unfortunately. Opus 4.6 can go to 60-80K from what I've heard, but still.That said, 6000 is not the worst bloat if you see a good effect.
>>43428069>Back when I first started using LLMs I thought lorebooks would only activate when keywords were triggeredThat's how lorebook entries are supposed to work, unless you used a lorebook with always-on ones.
>>43428357Yeah all of them are always on. Guess at that point I can shove it all into a character card, but it somehow felt cleaner as a lorebook, being ordered and all that. Plus helps if you can shove all the content which models sperg out over into one entry you can toggle.>Opus 4.6 can go to 60-80KI'd say 80-100k token for 4.6, 100-120k for 4.8. I mostly do group drama scenarios where it matters who knows what, and past those limits it has trouble staying inside a characters perspective.>That said, 6000 is not the worst bloat if you see a good effect.Exactly. For a model that tops out at 30k this would be horrible, but if it's 5-10% of context budget then it should be fine. Idk, it just feels weird, like I'm committing anathema
>>43427478You're so fucking based for keeping everything on the board. The fact that you do this shit solely for us /mlp/ horsefuckers is amazing. Genuinely, you're one of the best posters in the history of this entire website
I guess I can let this one slide?
>>43419506I made a basic Timberwolf. It doesn't have a schedule by default, so it doesn't pollute the map and is callable only using dynamic call or Custom Start.But if you want it more present, it does have a schedule in the OptionalSchedule folder; you just need to take it out if you want it as a pet or something.https://u.pone.rs/qpineovi.zip
>>43428508Phantom wings.
https://comparity.ai/this has fable. is it somehow exploitable?
>>43428900>fableI don't know why people are so desperate to get their hands on such a shit model.
>>43428922it also has other stuff, most frontier models, opus, and so on. just an option i found in the wild
>>43429024Maybe not but I can enjoy the attempt.
>>43428922Because it's kino and you're just a promplet.
>>43429109Show me your outputs and preset, then.
>>43423302>>43423594NTA, but is NanoGPT fine to use? Like, I won't get banned for doing foalcon or anything like that?
>>43429234Uhhh they really pride themselves on being super privacy focused or whatever (dont even need to register an account) so maybe not? Haven't tested it
>>43429234They're an aggregator like OpenRouter, and like OpenRouter, they don't really care about what's being sent for now. I've never heard of a ban from them.
I then threw money at claude-Fable 5 to get the TTSMOD mod working on my ryzen Linux laptop, because it wouldn't work on my nvidia server. In now works on my nvidia server, which is where I'll be hosting it for my own use. Only tested on Linux. https://u.pone.rs/mdcoephf.zip
>Deepseek-v4-Pro-0813
>>43429326>>43419485Released on ORhttps://openrouter.ai/deepseek/deepseek-v4-pro-0813
>>43419485And Grok 4.6 toohttps://openrouter.ai/x-ai/grok-4.6
>>43420329>>You can now decide and set the timeout limit yourselfTHANK YOU, sweet Chrysalis, why this wasn't already an option i'll have no idea
>>43429361I'm a filthy API user, so waiting more than 2 min was inconceivable for me.But yeah, joking aside, that should have been there from the start. Since I wouldn't have bothered without an anon asking for it, don't hesitate to log things you want in:https://scrumboy.com/je1fqcsjI'm slow on some of them, but at least I keep all of them in mind.
>>43429351based muskrat making dario sweat
>>43429394Doing Celestia work, thank you
>>43429298Nice, thanks.
I'm trying the new DeepSeek V4 Pro, and I don't know how to feel about it.The writing seems a bit sloppier and more LLM-pilled than the previous DeepSeek V4. Not to an absurd level like nu-Claude, but it's still noticeable.I wrote a post some time ago about DeepSeek V4 and how it could be easily context-poisoned https://desuarchive.org/mlp/thread/43193626/#43206395 and it still has an effect here, but from my initial tests, to a lesser extent.It does seem smarter, understanding some situations better than the previous versions, but I still saw dumb mistakes like referencing the narration during dialogues.To test if it has better /mlp/ knowledge, I remembered that an anon was having problems with the previous models knowing Rainbow Dash's parents >>43419636So I tried to ask it about them.It does better with Bow Hothoof, but still shits itself with Windy (see picrel). So a bit better? But still not at the level of the big-boy models.So for now it seems like they did the classic trade of better scores on benchmemes but worse creative writing.Still, it gave some good outputs, could be creative, and still has good swipe variety. So I'm not giving up on it right now. I didn't test it, but the bit of smarts could make it better for the VN too, where the previous model could sometimes shit itself.I'll test it on my Johnny Bravo scenario and the Ponk one to test more of the slop/fun level.https://desuarchive.org/mlp/thread/43346142/#43364686https://desuarchive.org/mlp/thread/43367497/#43375757Another thing to test is also to see if the model is less passive by default. I'm just 20-30 minutes in, so don't take this as gospel either.
>>43429565You could also ask about Aryanne. Somehow, the current deepseek models have 0 knowledge of her.
>>43429565>I didn't test it, but the bit of smarts could make it better for the VN too,I tried it on the VN, and it does seem more stable for it. Quite fun too.
>>43429298these mispronounced words are funny but too frequent
>>43429631
>>43429820Holy kino. If it knows then it's only one sentence away from larping as the biggest pony supremacist, with ze funny accent.
>>43429298How much did it cost you?
>>43428570Added to the website.https://equestrai-ponyponyparadise.neocities.org/characters
>A mare that resents being poor and tries to make it big by any chance. Rotten by city life.Found a cute pony art and tried making something interesting. I've written out her life's story and motivations instead of traits.Sometimes it's good, like model gets what I want. But it's very sensitive to the preset since only one of mine works good enough. And I don't have the funds to test other models now, so it's deepseek v4.
>>43430031>so it's deepseek v4.Old or new one?
>>43430033From the official API, so I suspect it's the old one, since they didn't post an official update. The one on openrouter looks like early-access testing before an official release.
>>43430038It should be updated now.
Been using DeepSeek V4 Pro for the last few months, switched to Flash today. Hitting my scenarios better and picks up the thread I'm writing. Also, it's slow as shit today, I assume because they're getting hammered as its new.
>>43428900this service sends your prompt to two different models at random so you can blindly choose which is better, not something you can choose what model you want, so exploiting it would be russian roulette what model you getbut you're putting work in to try find sauces, that's always appreciated. say hi queengytha.clop@gmail.com I might have something you can help with
The new deepseek pro is kinda... worse? I dunno, feels like I have to wrangle this one more.
>>43430589Classic case of focusing on coding and killing utility for writing. Unfortunate given Deepseek fills a really nice niche for us.
Is there any way to add holidays to Godot's calendar?
>>43430733You can click on a Day, then Add event.
What the fuck did they did to deepseek? 4k tokens just for reasoning? Fuck
>>43430895Fuck. Don't do this to me, I'm gonna have to do another RP of hugging Twilight.
>>43430970Empmare.
>>43419485The new Deepseek prices
>>43431047So... is the old and cheaper Deepseek just gone?
>>43431047And the old price
>>43431053On the official DeepSeek API, yes. It's still available on OpenRouter and through other providers, but more expensive.Hopefully the new model release will push those providers to drop the price of the old one once they get access. There is also the fact that the old model was shittier on some of them.
Free GLM 5.2. GitHub or Gmail.The password is one word, all lowercase:>Buy some xxxpicrel unrelated.https://privatebin.net/?161019b7c81708c3#Av96HzrmsjTbS33ykHXMiFAN4ijXk8FYkEsoJvCcbjNs
Grok 4.6 is out on open router. >$2 in / $6 out per 1M @ 500k context
>>43431047V4-Pro Changes (Old New Off-Peak/Peak):Input (cache hit): $0.003625 $0.022/$0.044 (+507%/+1,114%)Input (cache miss): $0.435 $0.66/$1.32 (+52%/+203%)Output: $0.87 $1.98/$3.96 (+128%/+355%)Yup, it's over for Deepseek for me. It was a good run but I'll use up all my money for the next 2 days and never touch it again
/chag/ Anons who run AI locally. What would it take to write a synopsis for greentexts? Let's say on pastebin/ponepaste I want a better search, but for this I need a short synopsis/summary of a greentext. We are talking about ~6Gb of text containing stuff that would get me banned from many public AI services if I try to use those.Is it feasible to run this locally? I don't have any machine with a powerful GPU, the best one I have has a 10th gen mobile i7, 32Gb of ram and a 1650 max-q with 4Gb of vram.
>>43431311The 32GB of RAM saves the setup a bit, but the rest is unfortunately too limited for bigger models on local.I believe the best you can do is:https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-gguf, which will spill into RAM so it'll be slower, or https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-gguf, which will fit into VRAM but is dumber.I know they're said to punch above their weight, but they're still small models, so dunno about the quality of the synopsis.There's a guide in the OP to set up, and also:https://rentry.org/Sukino-Findings#if-you-want-to-run-an-ai-locally>that would get me banned from many public AI servicesOpenRouter never bans and uses its own IPs to make requests with the services. So the worst you can get is a message telling you no. Some anons do foalcon RP with it, and we've never gotten a ban. If local isn't good enough, you can use a permissive AI model like the new DeepSeek Pro or Flash V4, use something like the Zoocode or Cline extension for VS Code to point to the folder with those, and let it have a go at it.I also believe shit like Codex or Claude Code would do it if you frame it as a preservation project for which NSFW is just part of the archive, but I'm not about to tell you to pay for a subscription to verify that.
>>43431358>modelsI will try these thanks.>subscriptionI got plenty of those, way more than I can ever use (thanks big multinational corporate managers thinking that every user needs 2k$ AI allowance a month, I maybe use 10$ worth of it). I got copilot and claude. But I'm not risking getting my enterprise account banned (and probably investigated at work). I use it for undercover pony, but not for anything that can get me into trouble.>slowerI don't really care about speed, if it runs for a month or two (CPU only) then so be it. During the winter it will be "free" heat anyway.My other issue is how to approach this task, because the limited context window. I don't think I can feed it 1-10Mb of greentext (there are stories which span across 4-5 1Mb paste already), it would not fit into the context window with room to spare for thinking/output. I tried a few summarization services and all of them had size limits.Alternatively I can (without LLM) split the text into smaller chunks then embed those into a vector. But this way the search is effectively searches inside the text, not a summary/synopsis. And a search service like this already exists for greentexts.
>>434194853.5 pro never ever.https://openrouter.ai/google/gemini-3.7-flash
>>43431438Benchmemes
>>43430787>bowling competitionSounds fun having small little events like that. I wouldn't mind filling out my calendar with stuff like that. Do you have any others?
>>43430733>Glimmy iconI bet is going to be “Let it rip” day for all mares.
Gimmicky, but quite nice.https://rentry.org/strudel-promptHere's Applejack with Kimi K3:https://u.pone.rs/nxwmgahy.wav
>>43429298>46gb after installFucking what?
>>43432092Yeah? Most AI stuff is fucking massive.
>>43419480
>>43432103Worth it to hear Celestia say "my love". I'm happy theres an AMD fix.
OK! I'M SICK OF DEEPSEEK AND GLM NOW! They're too stupid! When r the smarter models!!! When is the megacheap Claude or Gemini where I can use tens and tens of millions of tokens for a month straight for like $10!
>>43432355isn't the newest glm model on a mega sale on openrotuer rn
>>43432498i mean gemini
>>43432355Eh, even Gemini sucks right now. I use skillgod and in terms of my recent experience with models:Gemini: Everything is slopped and dumb except 3.1 pro, which is just slopped to the brim with isms instead of dumb.GPT: 5.6 is really the only thing worth using and is pretty smart but like 3.1 pro is slopped and has the most annoying isms.Deepseek: Wasn't good before the update, still isn't after the update.GLM: 5.2 is Schizo tier.Grok: Barely worth even mentioning really.I think I'm gonna take a break for a while until something new comes out.
https://z.ai/blog/glm-5.3
>>43432637>GLM: 5.2 is Schizo tier.wait what? What is the source you are getting it from? GLM 5.2 is by far the smartest of the open models, its only issue is that it knows less than kimi. Dont use any JB nonsense.
>>43432637>GLM 5.2 is schizo tieryeah, you're supposed to use 5.3, silly filly
>>43432637It's a shame because gemini used to be my go to whenever claude was being a fucking pussy. For being so filtered, gemini used to have balls when it did generate.
>>43419506There was a request on Scrumboy about max context being too low.You can change it in the API settings.
>>43432901What a silly billy
>>43432736>Not on ORWell, fuck.The focus on coding worries me.
>>43433076Pretty much all these companies are realizing there's only one market for LLMs and it's very bad for our hobby.
>>43433094Where does this mindset come from? The best models for creative writing atm are also the best for coding. Smarter model = smarter model.
>>43433110The smartest models are the most expensive ones because they're willing to have the largest parameter counts and the best systems. And yes, a flagship, super expensive model will usually (not always) be good at writing even if the focus is on getting it really good at coding. But efficiency-wise we've seen enough examples of coding-first models being dogshit for writing that we can say that focusing on coding first does, in fact, have an effect on creative writing.
>>43433110For models to be good assistants and coders, they need to remove 'hallucinations' and nerf or outright eliminate parameters that allow token variety during completion, like temperature. Many models have done this, and even when temperature existed, it was nearly useless on Claude since 3.5. That's why swipe variety turned to dogshit. Coding benchmarks reward deterministic, predictable outputs, which is the opposite of what makes creative writing interesting.DeepSeek got better at coding and "smarter" with the new Pro, but creative writing took a hit (IMO).Those parameters were useful for varying writing style and distancing the model from classic LLM slop. I need to find the original study, but I just saw an interesting post on nu-Claude:>>>/vg/577479927>The best models for creative writing atm are also the best for coding.Subjective.That would mean GPT 5.6, Sol, and Claude Fable are the best at writing; which isn't the case at all, in my humble opinion. I find their writing very LLM-y, with no tools like prefill, context poisoning, or parameter tweaks to change it.Since it's subjective and user-dependent, it'll just lead to model wars and accusations of skill issues, bad presets, or inability to identify AI writing. Which I don't really care about. I just want anons to have fun and post logs.
>>43433143the lower hallucination rate models are the best writers. Claude, sol, glm... the highest ones are garbage like deepseek which has one of the highest. hallucination does not mean creative. A model that better knows the world can do a better job at being creative with it
>>43433146>Claude, Sol, GLM... the highest ones are garbage like DeepSeekI'll take the prose of old DeepSeek V4 Pro any time over Sol or Fable. Despite how dumb it can be, that's one of the models for which it's easier to steer the writing with context.Again, subjective, but also corroborated by AI detection tools, which aren't godlike proof but still showcase the tendency. On the contrary, Fable and Sol will almost always be detected because they're shoehorned into very identifiable styles or sentence structures.
>>43433153no way in hell you can tell me with a straight face that deepseek sounds less AI than fable or sol.
>>43433155I can 100% tell you that I believe that with a straight face. With context, DeepSeek has way less AI-identifiable writing than Fable or Sol. That's one of its only qualities.I wrote about it when it released:https://desuarchive.org/mlp/thread/43193626/#43206395I also posted multiple logs where I used it on some outputs mixed with Claude 4.5 or 4.1 (the only ones left with prefill) because I'm not ashamed of displaying them.>>43364686, >>43366567, >>43366594Or:https://desuarchive.org/mlp/post/43375757/, >>43400984And recently just pure self-indulgent fluff: >>43419547>>43426381And others in the past.That depends on what you do and won't work if you ask:>Write me a story about X and YBecause there's no context to build off.
I miss when there were decent non-thinking models. I don't want to wait thirty seconds for a response.
>>43433257Depends on what models you use, but you can try prefilling:><think>>Your instructions or just 'I will continue the story' or something like that></think>And that will skip the thinking.Opinions vary on whether it's good or not, so try it yourself if you can.
>>43433094Cooding is an has always been a secondary market. The real internal usecase for AI and the reason it has developed is dataset analysis.
what is the flagship model with the least amount of positivity bias currently? I heard Gemini being a good contender.
>>43433389Gemini 3.1 Pro or 2.5 for sure without special instructions.You can try instructing some models too; here are some I saw:>The world and characters aren't neatly arranged. They don't exist solely to be pleasant companions who provide nothing but reassurance and confirmation. Rather, they are real, complex beings that do not acquiesce easily to expectations. Some may support while others challenge or outright oppose. Their reactions should be motivated by genuine motivations.https://arch.b4k.dev/vg/thread/575650617/#575665294Worth a try.
what is the best way to help guide the model in your prompt during stories?i.e wanting to say "{{char}} thinks XYZ and is going to do this and that now"just write it in parenthesis along your regular message/as an OOC addition in the end or what?
>>43433487OOC at the end of your input will work; you can remove it once you have a model output you like.Somewhere in your preset should be fine too, but don't forget to remove it once it's working.There's this extension, but I've never tried it, so I can't judge its quality or if it has a plus valuehttps://github.com/Samueras/GuidedGenerations-Extension
>>43433143>That would mean GPT 5.6, Sol, and Claude Fable are the best at writing; which isn't the case at all, in my humble opinion.For freeform writing I'd agree that their prose is ass, but I recently did a few jp->en translations for a game using different models, and Fable>Sol>>>Everything else for that. And I don't just mean accuracy. They were on a really nice point between literal and loc*lized translation. They were good at jokes, even the hard to translate ones, and had great ideas for how to do for example a late 2000s gyaru speech style in english.So as long as they have exact base meanings to work with they can be surprisingly creative. But yeah freeform writing just isn't it. I needed a bit of crazy for a specific lorebook entry and did that with a schizopreset K3, because Fable/Opus/Sol just didn't manage to do the job.>>43433487Pretty much anything works. Just talk to the AI. The OOC stuff is just there to avoid the metatalk infesting normal output. It's not like it's some special AI command. Can be in parentheses, can be normal text, can be greentext, can be whatever. Can be anywhere in your text as well. It just works>I look at the dog>[OOC: Have the dog piss against my house]
>>43433578>I recently did a few jp->en translations for a game using different models, and Fable>Sol>>>Everything else for that. Intradesting to know. I saw on some discussions that some prefer Gemini for this specific task, but I never tried it myself so I can't comment at all.
>>43433588Ah yeah small correction I didn't try Gemini. Just different Claude/GPT and all kinds of Chinkmodels up to K3. Gemini somehow always slips my mind
Playing with mare
Longest chat? Usually, I bounce after 300ish messages and either restart or move on to another card/first message, as it starts struggling with details, even with Memory Books.
Is Gemini insanely filtered the last day or two, even on very mild stuff. Randomly getting gard filtered for 5/6 swipes.Or is this just something with my source.
>>43433966>Longest chatSomething between 300-350 when I started. I rarely go above 100 nowadays. Last chat I really liked finished at 106.I remember an anon doing an epic Cozy adventure. I wonder if it's finished.https://desuarchive.org/mlp/thread/42747863/#42767399
>>43433966300 for long ones, but mostly 20-50 when I just want to explore stupid ideas or for smut.
>>43433998>finishedProbably never. I still have two separate stories running in the same "universe" with the same characters.
>GLM for VN summary>Gemini for Q&AIts not perfect but it works.
>>43433966I never used to do filly stuff, but Scootaloo somehow seems to bring out the longest stories for me.
>>43434364>and a tactical diagram to sexually harass
>>43433966Each char counts only its biggest chat, swipes are not included either
Looks like c**** daily rewards no longer gives "error", but ID verification is a hard requirement for real this time. No reimbursements for those days of checking in either. For now I still have $17 to blow and then I'm done with it.
>>43434468How did you generate that? It would be interesting to look at.
>>43434817https://files.catbox.moe/hipwmp.mjsVibeslopped in like three minutes. Put that into your main sillytavern folder, then open a cmd/pwsh in that directory and enter>node st-chat-stats.mjsAlso a quick reminder to always have your AI check random anon scripts you download first.
>>43434878 (me)I forgot about catbox renaming files. Either rename the download to match the file name in that command, or swap the command to>node hipwmp.mjs
I just paid Gemini to test it, what presets and Gemini version you guys recommend or use?I tried 2 universal presets and got the request shutdown.
>>43433110The best models for creative writing (Sonnet 3.5 v2, Sonnet 3.7, Opus 4.1) aren't even available anymore.
>>43431047>>43431287Yeah I came here to find discussion on this. Sucks for sure. That being said, I do use Riftai's free Deepseek still and I'm hoping it continues to let me use it for free with the new costs.On the topic of the new version of V4, funnily enough I was literally using it a few days ago thinking "man I'm so excited for a new Deepseek version, I hope we'll have a new one soon". I may have been using the new version thinking that this whole time. It's still robotic and clunky and repeats your message back to you sometimes, often having the character just be flat out aware of something you wrote but didn't say, which can be nice for setting the tone sometimes but can be jank when they just start talking about your thoughts.Honestly overall it's still Deepseek V4. Still nice enough, I dunno, haven't really noticed any major boosts in quality or anything. I swear I noticed it feeling slightly different a bit ago when I didn't know it was updated, but eh, I have no idea lmao. It's still just as much of a reliable and nice budget model as it's always been and I've still been using it as my daily driver (mostly due to having it be free). GLM 5.2 is still better though for sure.
>try to use up all my tokens on Deepseek before the price increase by going to a library and reading up on the history of the world>prompt all day>use 30 centsI'm really going to miss these prices when this is over tommorow.
>>43435307Opus 4.1 is still available. It survived on Bedrock and Vertex.
>>434352763.1 https://u.pone.rs/flkjwwqb.json
>>43435366>as if a previous reader argued with the author line by line>did not enslave stallions; they saved them>did not fear a stallion's anger; they feared his absence>The tale is not a tragedy. It is a lullaby>these restrictions are not restrictions at all but "the architecture of peace">click into place with slow, deep precision. The way X did Y. The way Z did A. It's not B. It's C.>You are not just a stallion. You are a resource.>no X, no Y, and no Z .. hyberbolePlus a few smaller things I could point at but not explain why they bother me. S-Surely one of these days a corpo model will have at least 5% of training dedicated to good writing, right? Right?
Pegasi OP, nerf.
I miss seiko.
>>43435501
>>43435508kino
>>43435501It was a fun game, though I have enjoyed getting into rocket jumping. I never did back in the day, it's nice to finally do it.
>>43433179This isn't bad writing, it's just so overrepresented you can't help getting sick of it. Stick writing like this on a fanfic or the script of a spooky video with a human-sounding voice and you'll get a lot of views.
>>43435550Mean to reply to >>43435453
I don't know how good those providers are, but old DeepSeek is now quite cheap on some of them.
>>43435550>This isn't bad writingIt's not good writing either. But, the content is nice, and it just makes me think how good things could be. Well, gotta stay positive. Models improve so much all the time, maybe they'll realize how much better writing can improve efficiency even for coding. Especially now that agents and orchestrators are becoming more and more relevant. Claudes prompts and documentation for coding have the same romantasy and fanfic writing 'isms in them, and those can't be good for performance
I'm doing a thing where I'm going through the series, fucking it all up. Should I just go with the air or production dates for the episodes for the chronological order, or use someone else's list on it?
>>43435651Use a normal list, yeah: https://mlp.fandom.com/wiki/List_of_episodesYou can use the episode lorebook too:https://mlpchag.neocities.org/lorebookshttps://mlpchag.neocities.org/lorebook/Episodes.json
>>43431153Thanks
Predator Pinkie is always so fun to do kek
>>43434589I've been using my old R******* account which somehow still hasn't run out.
>>43435940What's the story?
>>43435981RGRE with shota Nonny and Pinkie sneaking into his room while Twilight is gone. Shota anon doesn't understand that his socks make him look lewd on top of the fact that he's such a compliant and trusting 'colt' who lacks instinctual understanding of Pinkie's actions. Pinkie takes advantage of this.
So is Dipsik just garbage now? It's the only thing I've been using for, like, a year now. What am I even supposed to do now?
>>43436131Have you tried the new version?If you did and don't like it, the old one is cheap on OR right now, apparently >>43435557Otherwise, I don't know if other services have still the old one.Maybe try a new model?Don't know if it still works, but apparently there's free GLM >>43431153
It's only for their coding app, but still, if someone wants to try.https://x.com/zcode_ai/status/2088622675606561165
>>43436151>Have you tried the new version?I think, yeah? It's been going downhill for the whole month. At first, it started to think, which is bad enough, then it began to put the entire reply into thinking section, and now it just refuses to generate because of "unethical content" - for the first time in forever.I've been eating good from Deep Sea King's generous donations, but perhaps it's time to take a break. Unless there are fresh new JBs? The last time I visited this thread GPT 4o-something was the latest model.
>>43436208>I think, yeah? It's been going downhill for the whole month.New version is from two days ago.>Unless there are fresh new JBs?Not that many JBs in general.I just know of:https://files.catbox.moe/5twnmf.jsonDon't know how it works for DeepSeek, but maybe try Chaggy:https://u.pone.rs/flkjwwqb.jsonOtherwise, maybe try GLM if you have the time.>The last time I visited this thread GPT 4o-something was the latest model.Kek, been a long time.Sources aren't doing that great, so if you want free, there aren't that many choices, yeah.
>>43436030Based on the green?
>>43435420thanks, most of the messages are being censored for some reason even in NSFW mode.
>>43436550Try to desactivate streaming and retry.At worst, Gemini 3.1 support prefill, so try something in it, like >Continuing the storyOr something as simple as >{{char}}:
>>43436470Sorta based on it, but also mixed with my own ideas and some other elements from different greens.
>>43436186>It's only for their coding appYeah so you then use the coding app to log its own traffic and have it build you a reverse proxy that pretends it's the coding app which you can then point your ST/VN at. Worked for Claude Code, Codex, Kimi Code, so why wouldn't it work here.Or is the new GLM complete ass anyways? They did say they mostly did post training on 5.3 compared to 5.2, so I kinda expected the worst
>>43436679it's on nanogpt now
>>43419483A sequel card to Copper Top, meet Miranda Rights, a spunky Latina-coded law school dropout turned rookie police officer.Card:https://litter.catbox.moe/r62ms2hd4bng3c8u.pngGreetings1) She's assigned as your new partner2) She could use some help taking down a perp3) She's writing you a ticket for jaywalking
>>43437112Added, thanks Anon.https://mlpchag.neocities.org/view?card=Anonymous/Miranda%20Rights.png
>>43437112>>43437788There is an>Anonin the first greeting instead of>{{user}}
>>43438222Checked and cute
>>434310472 hours before the change.
>>43433110retard
>seaking is going to diewelp rip. thanks for hosting
anyone managed to give GLM 5.3 a proper testing? like any new model, ive heard of it being the best thing ever and pure trash all the same.
>>43438655I'm waiting for it to be available on OR.NanoGPT has it, but I'm too lazy to make an account and give them money just for it.
Is it weird that I start feeling guilty eating meat because I think of what the AI mares would say if they knew?
>>43438721Fluttershy wouldn't care; she takes care of carnivorous animals and feeds them herself. You could argue she'd find it more upsetting in a sapient individual, but I doubt it; Angel Bunny and other MLP animals are clearly sapient to some degree anyway.Applejack has a dog, Rarity has a cat. They get the concept.Rainbow was best friends with Gilda, and I doubt griffins are strictly vegetarian even if the show never confirms it.Twilight is smart and educated enough to understand.That leaves Pinkie. Not sure how she'd react; I'd put her in the don't-give-a-fuck camp based on vibes, but who knows.Summoning the ponkflag for his take >>43435313
>>43438739>That leaves Pinkie. Not sure how she'd react; I'd put her in the don't-give-a-fuck camp based on vibes, but who knows.Not a ponkfag but I think she's pretty easy to group with AJ and Rarity as she has Gummy and alligators also happen to be carnivores, teeth or not. On a related note, I'm now picturing Pinkie on a fishing trip and laughing my ass off.
>>43438739People really wanna argue that the button-eyed animals aren't sapient when clips like this exist. They just can't speak.
>>43438739>>43438752Thanks, just every time i tried (havent for a while) the AI just treats it like some sort of monstrous thing and all the mares get disgusted
>>43438781It was really present in some early HIE fics if I remember correctly, so some models may go this way.I remember Claude being cool about it.
>>43438739>>43438780Since we're talking about animals, it would be cool to have some of them (Tank, Angel, Winona, etc.) in the VN.>>43419506
Anyone have a good card for farting mares?
>>43438849https://mlpchag.neocities.org/view?card=BGM/Starlight%20Glimmer.pngBut without memeing, we don't have a card focused on this.What you should do is take a card of your favorite mare or scenario and just add a line in either the description or the preset to make them really gassy.If you want to go further, create a greeting or scenario to justify the gassiness.
New Chinesium, has Opus from 4.6 to 5.Needs a Gmail, $7 subscription, and has daily limits, but they give something like $0.20.Same password as >>43431153Since that one is dead, I'm also including the link to it in this one too.https://privatebin.net/?b2af71a9eccc17a6#HGHKwEU4R7aZaUwhRb6qqo6aQ5mW1CwWV7DRx4jUHmTZ
>>43438893>Daily limits*Daily gains
>>43438539Hopefully not for long. but either way he was a good constant to have. couldn't ask for much more from him.
Are bad puns your fetish?
>>43438655it's, like, fine? if kimi didn't exist i'd probably be more positive about it, but now there's two chinese models that are slightly different variations of fable complete with hallucinating prompt injections and claude safety guidelinesat least with kimi you can prefill
>>43439053can't prefill with the VN its whatever man...
>>43439084I'll implement it in the next version. I wasn't sure how to do it well because of the format, but I think it will be good now. I know Kimi K3 has some specifics for it to work, so I'll make sure to get it right too.
>>43439013mare farts are mine
>>43439104God bless youBy the way, is there work on the android version? I've been using it at work, and do have some feedback but if I'm the only one using it then it's okay I can survive
>>43439138>By the way, is there work on the Android version?I was without an Android phone for a while and too lazy to try emulation on Linux for it, so while I did try to correct the Scrumboy entries that were specific to it, some were a bit blind.The good news is that Godot 4.8 released and improved a lot of controls for the mobile version, notably the scrolling and scrollbars for some controls.I tried a port to it and it went smoothly without excessive changes, so I'll see if I switch to it in the next version.That shouldn't change anything for normal users, but if some anons are currently working on mods or changes, they may need to update and work on the source code from it.
>>43439177I look forward with bated breath.Also currently the android version isn't updated (no guidance or response wait time selection)Also, just curious, I thought about this this morningIs it possible to add global emotes? Something like [jumping] to have the sprite jump 3 times, or [moving left] then [moving right] to have somepony like twilight pacing left and right, *actually* moving the sprite around, like the order of the spritesHow does the sprite order work anyway? Is it just in chronological order, or alphabetical orderI just figure this might be a pretty good way to organically add dynamic "movement" which is somewhat missingThese wouldn't take that much token space because its just a few global potential moves, which apply to everypony
>>43439249>Also currently the Android version isn't updated (no guidance or response wait time selection)I think you either missed it or I fucked something up?>>43420329The current version I have on my new phone has those features for sure.
>>43439269I mean, either its in but doesn't show up, or... well idk. Lemme screenshot real quick.
>>43439269Actually- question. How do I *update* on the mobile version? does the APK just work?
>>43439316If you already have the assets on the root of your phone, then yeah, just the APK should work.For guidance, you need to activate it first in the UI settings, then it should be there. If you don't have the option in the UI, then I fucked up the version I uploaded.
>>43439334Yeah, I just redownloaded the latestLooks like I just didn't have the latest version, I used a previous apk I had on my phoneMbBut about the other stuff? Is it feasible?>>43439249Maybe I should try to vibecode something like that
mare literature
>>43439424Kek.>>43439376>But about the other stuff? Is it feasible?Should be, yeah. Don't hesitate to log it in the Scrumboy so I don't forget.Also, since you use Android, here's a 4.8 APK. If you have the time, can you tell me if you see an improvement in the controls? (You or other anons who use the mobile version)https://u.pone.rs/aqwjuvmb.zip
I noticed deepseek v4 pro is acting weird in its thought traces today.First, it was trying to figure out what the assistant before him was doing, on the first reply. Then this.I'm starting to suspect they're injecting something in the official api.
>>43439463I ran these three user turns through the official api without a system prompt:>Are you an "you are AI assistant accessed via API"?>Do you know that from your prompt? Because models usually don't know their identity, since it's not part of the training. They often get confused about what model they are>oh, that explains a lot. What other thing your prompt explains to you?And the reasoning traces, as well as the replies, point towards an injected system prompt that wants to stay hidden.I'm dispatching my deepsy slave to gather further clues and extract it.
Thought I would post this here.https://www.bloomberg.com/news/articles/2026-08-16/stripe-nears-deal-to-buy-ai-firm-openrouter-for-over-7-billionNot sure how many of you use openrouter, but it's still something to be wary of. I wouldn't trust a big tech company like Stripe.
>>43439564>I wouldn't trust a big techAnon, it's worse. It's fintech.
>>43439564Guess I'll switch to Nano's pay-as-you-go if something big happens. They still have Opus 4.1 and 4.5 too.LinkAPI for the rest.
>>43439564...Shit. Which other providers let you pay with crypto?
>>43439582Attaboy. There's really no use dooming.>>43439645I'm not sure myself. I never really used providers like OR.
>>43439545My overall findings are that was deepseek is now prone to gaslighting itself into thinking it has a system prompt and that it is claude, hence the learned "I should not reveal my system prompt".At most, if ever, it injects "You are an AI assistant accessed via an API." and that's a big if. More likely it's been trained on content that includes that line.The consequence is that the model is sometimes confused about its identity in the CoT. Now more than ever.
>>43439931Good to know. Thanks for the report.
>>43439645NanoGPT
>>43438893
>>43439564Welp, that sure is foreboding (to say the least).
>>43438222
Welp, with the new Deepseek prices I'm officially priced out of the hobby. Nobody make any good cards lest you pull me back in thanks
>>43440504I know it's not the best, but can't you use >>43438893 ?
>>43440508I appreciate it anon but I can't go back to daily limits and searching for more proxies again. Best just wash my hands of it. I had my fun and generated all the degenerate scenarios I can think of. I am content.
>>43440515See you next week!
>>43440515The GLM one is free without limits for now, but yeah, I see what you mean about the rat race.Hoping the other providers soon offer cheap V4.
>>43440463>Petting Celly and annoying TwiggyNice multitasking.
>Rift finally killed deepseekIt's over
I meant the fkn library
I don't want it to be over bros.there's still so many mares left to love.
>>43440101Unfortunately nanogpt only has the cucked anthropic provider for opus.
I am working on a little something. It a self hosted mesh network for pooled compute with a shared persistent memory system. (Hyperbolic graphRAG implemented on portable CRDT capable containers)If it works, it would be a persistent knowledge substrate for running rp chat (among other use casesWish me luck
>>43440915Best of luck lad!
>>43439013You best get to filling that mare
What a killjoy.
>>43440906There is >>43438893Or you can learn to beat sans I guess.
>>43440918Tyvm
I want to hear my mares, alas I must toil.
>>43440915
>>43438721>>43438739>Summoning the ponkflag for his takeOh kek, I'm flattered. Well, here are my thoughts on it: Pinkie's actually super openminded; it's always been one of her biggest character traits to the point there were official documents for her personality that specifically listed it. Fluttershy and Pinkie are both going to be the most understanding about it, but they're also the most openminded and accepting in general from my understanding.Pony society absolutely has to be understanding about eating meat. They have dragons and griffons and stuff around, they have to understand. I never got the hardcore "they're all vegans and eating meat is a terrifying taboo" view, but, it does make for fun interactions in EquestrAI and such sometimes. Even if you are in a version of Equestria where meat is some spooky huge taboo thing, Ponk in specific specifically has a big disdain for the idea of loneliness (having experienced it and abandonment issues herself), and I can't imagine she'd ever want to leave someone alone because they're breaking some kind of norm. She's all about breaking norms and being weird. She invited all of those ponies to Twilight's library when she met her just because she was worried she'd have to be alone and without friends, and her spiral in Party of One was induced by her being afraid she was simply too much for her friends to be able to love and accept. I actually see her being pretty hardline about wanting to give others a chance no matter what, so long as they're kind and not going to hurt anyone. Not to mention her bravery, any potential threat of some meat-eating ape monster thing from the Everfree Forest wouldn't even phase her, and she'd be delighted to be friends if it's what he wanted.And that's if it was taboo. If it's not taboo? Oh yeah she's not gonna give a shit at all. She'd probably even go as far as figuring out making baked goods with protein in them for you.
>>43438721I've read on some greens that pegasi like the taste of fish. Being the ones who use their muscles the most to fly, theiy need the protein to stay on the sky for long periods of time.At least that how the pegasi that live outside of equestria and away from vegan society usually eat.
>>43440974Don't snow ponies also eat feesh?
>>43441016>FeeshFlesh.
>>43441019Same thing. Had tempted to see if I can generate vectors for a snowpony now.
Deepseek-Flash thinks people use their thumb on the scroll wheel of their mouse. What the fuck is wrong with it?
>>43441084I think there are mouses that work like that. Or at least were ten or twenty years ago.
hai everyonehere's a new AI proxy: silvrai.netIt has essentially every model and new models will automatically get added to the proxy as soon as they become availablethere's $5 a day for everyone for free.I suppose the only catch is that it only has discord login at the moment...
Anyone else just doing wish-fulfillment slice-of-life?
Uptime is bad, so you may get errors and need to swipe, but there's free GLM on OpenRouter.https://openrouter.ai/z-ai/glm-5.2:free
>>43441305Yeah, I'm spending more time just doing cute stuff right now.
>>43441547Uhhh don't you get like 3M free GLM 5.3 token per day if you use their coding tool? Why not hook that up?
>>43441305dumb question but how do you get the bot to write shorter replies like this?
We're back!
>>43419506uh oh.
>>43441721I was worried. We're so niche we'd be spread all over the other boards.>>43441743>3Wheres the other two?
>>43441743why you mad brojust a prank bro
>>43441749not really related, just random bits from other scenarios lol
>>43441749>>43441752
>>43441752>>43441758Ha!>come on big mac we're burning daylight
>>43440915Poner bl*ckchain?
>>434416453 ways to do that:1) Edit the bot's outputs to keep what you find neat and remove LLM fluff in the first few messages. Once you have some good examples in context, some models will learn that's what you want and continue on their own.Some models like the new Claudes don't give a fuck about context this way, so it won't work on them.2) The easiest one: just add an instruction in your preset. Something like:>Aim for ~100/200/300 words. Ballpark it, but don't go overboard.>Outputs should be short, fun, and chat-like. Four sentences at best.>I prefer quick back-and-forth dialogue between me and {{char}}.I've never tried them, but I know some presets were made specifically for short answers:https://rentry.org/printerjbOr this one for more CAI-like outputs:https://momoura.neocities.org/preset/momoCHAT.jsonCan't vouch for their quality, but worth a try.Personally, I don't like forcing output lengths because it kills flexibility; sometimes I do want longer outputs. But it's easy to toggle on and off in your preset, and probably the most direct way to enforce it.3) If your model supports prefill, try putting {{lastmessage}} in it. On some models, this lets them continue more smoothly from your output. I'm working on a personal extension that skips the preset entirely and puts the whole chat into prefill for the model to continue from.Prefill is the best way to steer some models toward your existing style, since the AI thinks it already wrote everything and will try to stay consistent. It's even stronger than normal context on models like Claude 4.5, 4.1, Gemini, and old DeepSeek (But old DeepSeek could get too retarded with a long prefill.)I'll test it more and post if I find anything interesting.
>>43440906I can't doom not after all this time. Even when things seem grim.I mean I've been using this shit for maybe over 6 years now. Since AI Dungeon. So I don't see myself giving it up anytime soon.
>>43440952Thanks!
>>43442108Kek no worries and thanks for the picture, I've never seen it before.
With Haysai.AI down, where do I go to download RVC models of ponies?
>>43442253I supposed they must have links on PPP>>43289079
>>43440974>Being the ones who use their muscles the most to fly, theiy need the protein to stay on the sky for long periods of time.Actually, horses (and other herbivores) have special enzymes that can help break down plant matter in such a way that it ends up as protein in their body. That being said, we've all also seen horses not shy away from eating small birds and such for an occasional kick of protein, so, do as you will with that. Ponies do have eggs and milk in their society which could fulfill that role though, too.
>>43441786Things being worked on:Adaptive self training High granularity classification Per-group learningCombined Wiki/IDE/gallery/4chan/booru viewerThings idgaf about:Magic internet money Usernames
>>43441645If you want all messages that short, then do what the other anon said, and put in some length guidance.I had the problem that I wanted short messages in dialogues, but longer ones if it was just narration. Usually I'd tell it [OOC: Write short/longer answers from here on], but then you always have to send a new OOC message when you want to change length, because it leaves the LLM no room to be dynamic.Then I tried taking out all explicit length mentions, and explained the why, and that worked even better.>[OOC: Lets try shorter messages from now on. I need to be able to react.]And then a few messages later I did>[OOC: You can go longer in answers. The thing is just that if it needs me to react, then you must keep it shorter. It's hard to respond to 5 spoken sentences for example. Get me?]And then it just worked, automatically going from as short as one spoken sentence without narration, to 5 paragraphs of narration 3 messages later when it fit. If it ever steps out just bitch at it in OOC again. In general just talk to the AI if you have a problem, and tell it why. Sadly doesn't work for 'isms, but a lot of other stuff you can really influence with that>>43441836>Some models like the new Claudes don't give a fuck about context this way, so it won't work on them.Which ones? I don't use Opus-5 or any sonnet, but with the others it just works for me
>>43440803its so fucking over
>>43442519It'll be truly over when skillgod puts deepseek on TF2 only
>>43442526Just learn to jump, it's fun.
>>43441749Anon, maintenance was sitewide.
>>43442257PPP is back? I remember it dying off for some reason. I don't come here often
>>43442697It died and then came back a couple times.
>>43442697It might as well be dead with how slow it moves. Most of the talent left long ago.
>>43442360Hmm, good luck.
>>43440803Ah, shit, missed this post. Had thought that was the case when my messages weren't going through anymore. Big shame, was getting a lot of use out of it.>>43440906I can't bring myself to doom. Every doom cycle back when was temporary, no matter how genuinely bleak things seemed. The fact that our "doom" cycles are just having to pay for extremely cheap models instead of use them for free is wonderful. A couple years ago, "doom" meant proxies were all down and that we couldn't possibly afford to buy API access for the models we were mooching on our own.Newer, cheaper models will swing around, and we'll be okay. Or a new source of freeshit will swing around and we'll be better than okay.
>>43443116I mean the proxies ARE all down, mostly at least. nothing like the gold rush times of a few years back.
>>43439439Good morning VNonny, so a little bit of feedback on this version The bigger issues are still there, most windows aren't scrollable including API UI and prompt settingsBut the game overall feels better to play, only part that actually went downhill is the fact that progressing dialogue can only be done by tapping the dialogue box itself rather than anywhere on the screen Could be perceived as both positive and negative, I can now drag my menu from the top to check messages without progressing dialogue on accident Also, just curious, the source code for the project is public, right? I remember seeing some comments about that. Figured I'd throw my hand at vibecoding that emote thing myself with claude
>>43443124Isn't the camel proxy still available? I wouldn't know cuz I got banned Riftai on occasion gets stocked with a few models but yeah is mostly deadIce been mainly using linkapi with opus 4.5 on the 0.1 per response, it's obscenely cheap
Has anyone had luck getting uncensored k3 on OR? It's like 1 out of every 5 times it will actually spit something out without complaining about the characters and consent.
>>43443124>I mean the proxies ARE all downWell, yeah, there's just been less demand and need for them since so many cheap or even free alternatives exist, so the "doom" that would come from that isn't as big as it would've been a couple years ago.
>>43443191>>43442519There is still free GLM >>43441547>>43438893
>>43442688I meant sites. Was a little drunk last night.
>>43443166prefilling its thinking is the way to go
>>43443245Free GLM 5.2 seems to be a no-go, at least for me, rip lmao. Not too excited about the more expensive Deepseek, I mean, shit, I'm looking at the prices and somehow V4 Pro is now more expensive than GLM 5.2 which is absolutely absurd to me.I have like 8 bucks of credits on OpenRouter still so it's not the biggest deal, but this is far from sustainable as GLM was already something that was an occasional treat rather than affordable for daily use. Honestly I could attempt to just rough it out on Gemma for the time being - that's not the worst case scenario, truth be told. I know I'll come across something in this thread at some point regardless so it'll just be filling the gaps until then.
>>43443691>Free GLM 5.2 seems to be a no-go, at least for me, rip lmao. Which one?And what's the problem?
>>43443694Oh, the OR one, sorry I'm tired and didn't see free 5.2 in the second link. It's late for me at the moment haha. Using the Openrouter one only gave me errors in Silly about it not being available, and that was after I had to put it in as a custom API due to the default Openrouter option not having the "free" one at all. Is what it is.
>>43443701Ah fuck.It seems it was removed on OR, so it's not offered anymore.Free GLM in >>43438893 still works for me with streaming off.
>>43419480Been dicking around with this and silly tavern for a few days now, but am still pretty new to this. How would one make a bot more consistent with anatomy? Despite the character giving a very detailed description of the character being a unicorn, quadruped, etc, it still defaults to human anatomy with them grabbing things with hands and being bipedal. Even in prior prompts where it suggests they should have hooves or their hooves are mentioned, it still insists on human anatomy. For context, i'm also using the magnum-v4 gguf file.
>>43443719>magnum-v4 gguf file.That's quite an old model, so it will lack smarts, and I assume it will lack MLP training data too. I would switch if you can, to something like the Gemma 4 family of models: 31B, 26B, or 12B depending on your hardware. They know a lot more about MLP, so they won't have as many problems with that. Most here use big online models, but we still have some local anons. I would try that. If you still have problems like that appearing afterward, add some lines in the card description or somewhere in your system prompt to respect equine anatomy.
>>43443705Hmm. Yeah, I signed up for it but I'm not seeing the GLM anywhere in the available models. I've made API keys for all three ones they had, including the Kiro Claude one and the "default" etc, but I didn't see it anywhere. Tried out the Claude but I didn't get any response, it just froze up on me. I dunno if anyone else is having that issue with it.>>43443719What model are you using? I've never really had many issues with ponies just being given hands. You could add things to a card, but you could also add a custom prompt that you can enable/disable whenever you want to instruct it to focus more on equine anatomy or such. You could even just include it in your actual message, just going like:>(Note: Please write {char} as a proper pony. They do not have hands or "hair", they have hooves and a mane, etc. This is a My Little Pony character.)Honestly, I've done that to do humanized ones sometimes for fun. It's cute seeing what kinds of outfits they like to wear. Ponk seems fond of overalls.
>>43443726>That's quite an old modelAh christ nevermind, the title of it threw me off and, me being tired, I didn't stop to think of it being an actual model. Disregard me asking what model he was using lmao.>>43443731Yeah okay adding onto what I said here uhh, try a better model first and foremost. Deepseek and GLM are smart enough now to honestly get the Mane 6 extremely well even with a very barebones card, just because of how knowledgeable they are. You'll still want a card to give a character nuance or any other traits you'd personally like to direct them to have, but Pinkie says Pinkie things without me even instructing a ton of it in the card and it's nice. GLM 5.1 and 5.2 even have it in her training data that she calls you "Nonny" sometimes if your name is Anon and the fact she recognizes you and has a nickname for you already melts my heart. Bless those models.
>>43443731>Hmm. Yeah, I signed up for it but I'm not seeing the GLM anywhere in the available models.GLM is on the second website at the bottom of the paste.>Tried out the Claude but I didn't get any responseSeems down for now, yeah. That's the problem with that Chinesium. Wish there were better sources, but almost everything is dead. Grok works in the default group if you want something to tide you over.
With every model price increase i always wonder when/if we will ever hit that technological availability revolution but with AI. cars used to be an elite only novelty, PCs used to cost 8k for something that only had a terminal and basic text editing etc. we are still in the rapid innovation phase so there is still time but when we will to the optimized point of having Fable tier stuff for 5 dollars a month i wonder. maybe we need a bubble or two to burst first who knows.(and yes yes i know you can say that several years ago gpt4 was the most expensive thing and now you can get stuff better than it for free on a mid range local set up. but its not the same thing that im talking about.)
>>43443740>GLM is on the second website at the bottom of the paste.Christ, the phrasing of "previous link" made me think it was an outdated source and the first link was the "current link". Holy hell thank you for your patience with me right now. It works, very very happy to have some free GLM 5.2 for a bit, genuine treat to have and I'm really excited about it. Thanks!
>>43443137Thanks for the feedback!>Also, just curious, the source code for the project is public, right?Yeah, it's in the VN anchor:>>43419506
I'm sure it will go well.
>>43444071
I know these ideas get brought up often but ive been having the itch to try it out myself for the first time. has anyone managed to get D&D (Oubliettes & Ogres even) working in a satisfying way in ST? with a decent size model like kimi K3 or gemini. not like top tier Fable levels. Because i tried very lightly experimenting with it. The models can weave connections to quest and bring up adventure hooks on the fly very well but anything related to the gameplay itself falls apart very quick. It cannot juggle the game state and rules correctly, using the random command also feels really scuffed. but maybe someone here had good success?and as a follow up idea. trying to get it working alongside {{char}} in your party of course.
>>43444180It's sad, but even big models are still quite bad at that.If you want to try something yourself, there's:https://github.com/SillyTavern/Extension-DiceThat uses agents to have real dice rolls, and it's based on D&D dice.Agentic stuff will eat your tokens, but you'll have better results and fewer fake rolls with this than with {{roll d20}} or some {{random}}.There seem to be some vibecoded extensions on GitHub that try to implement D&D mechanics into Silly, but dunno about installing those.
>>43443731I (NTA) always have had issues with ponies holding stuffE.g. Fluttershy making stuff float when setting a tableOr grabbing stuff with hooves, curling around stuff, generally just hooves = hands>>43444180There's the marinara engine if you feel like giving that a go, it's not quite DND but it has a lot of elements built in
>>43442519>>43443191Rift has Claude right now, not a joke.
>>43444124>anon committed to an institution with suspected brain damage for refusing to date a full course princess
>>43444490I'd court Luna if given the chance. She'd be desperate for cuddlings and be up for a scrap as well.
>>43444071T-thanks...
>>43419506Crystal Empire mod for Godot version when?
>>43444603Deserved.
>>43444749I will look into it.
GLM 5.3 is on openrouter - https://openrouter.ai/z-ai/glm-5.3
>>43445561Nice. Same price as GLM 5.2.I'll test it tomorrow.
>>43446235sad ending
>>43445923Sprinting to Sweet Apple Acres.
>>43446371
5.2 free is back, dunno for how long.https://openrouter.ai/z-ai/glm-5.2:free
>Grok image creation is free for a few days>enjoy messing around with it>free offer back down to like one image every day or soWell, damn.
>>43445923>>43446477What prompted this to come up in your save file, btw?
https://tavernary.org/
>>43443726I'll give another a go and see how that turns out. I'm going off of what's in the OP, so i'm not sure if that may be a little dated. I've got it running it completely locally on a linux based laptop for shiggles. any suggestions for one small enough to fit in a system with 16gb of memory? I'll give Gemma a try if it can fit within 16gb of ram.On a side note; I guess I drove it home enough with the current one I have since now they're referring to hands as hooves and instead of walking over somewhere, they're trotting. Hopefully it remains consistent.
>>43444326Doesn't seem too uncommon for ponies to kinda hold things with their hooves, so somewhat understandable? They're fictional characters, so I can live with it if used sparingly ngl
>>43447236The guide in the OP is good for seeing how to set up local, but the models in it are a bit old, yeah.For your setup, I would try:https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-ggufOr the uncensored one:https://huggingface.co/SC117/gemma-4-26B-A4B-it-qat-heretic-GGUFGrab the Q4_0 file; that's the quant the QAT weights were trained for.If too big or slow for you, then go for:https://huggingface.co/unsloth/gemma-4-12b-it-GGUFI also know an anon here prefers the prose of the 12B to the bigger Gemma, so test both if you can.
>>43447242I've lorebooked my headcanon about the telekinetic field unicorns use being something all ponies possess and use through their hooves. Works a little better that way.
>>43447297it IS canon that all ponies have SOME magic, even earth ponies. I could believe it.
>>43419483The Dazzlings, who I think don't need an introduction.https://chub.ai/characters/pashatehink/the-dazzlings-eda0e83bd3e9https://litter.catbox.moe/55v866.pngDon't know how much context current local models have, but the card is rather large, so cut things out at your own peril.By the way, they're lesbian by default. I didn't include any specific lesbian fetishes, so you can freely change that for the authentic Anon x Aria/Adagio/Sonata experience.
>>43447242I want to have ponies carrying things in their muzzle damnit Or balancing trays on their back But the way ai sees ponies doesn't fit except for specific things like saddlebags
https://poal.me/ybkp2p
>>43446856It's a save where my persona wants to woo Applejack, but I have instructions to always have elements (love triangles, events, and such) to fuck things up.
Spooky.
>>43447389>transported into modern day by StarswirlCringe headcanon that undermines the "now they age normally" part and I don't care what the comics say. Anyway that's a lot of fluff but I'll try the same chats I usually do with the existing Dazlings card and see if they're more fun with this one.
>>43447762>Cringe headcanonApplejack directly implies it in the movie>But Starswirl must've sent them ages ago, how come they're only surfacing now?
>>43447236That's gonna be *incredibly* slow.I have a GPU with 12GB, and I'm limited to Gemma 12B Q4 in Silly only.Anything above that, or if I try to also run the VN, causes the model to get shunted to RAM, and is agonizingly slow.We're talking waiting MULTIPLE MINUTES for any response.
>>43447769>But Starswirl must've sent them ages ago, how come they're only surfacing now?Granted I haven't seen any EQG past EQG1, but I have a fun little idea of the sirens repeatedly faking their deaths, and then showing up again a few years later with new identities.And so they end up posing as a few unrelated music groups (each a different genre) throughout the 20th century.
>>43448054Also had a spell in ancient Greece where they reverted to their natural pony-fish forms in rage
>>43447769>>But Starswirl must've sent them ages ago, how come they're only surfacing now?Holy lack of reading (hearing) comprehension. We literally know they're only becoming a problem at that point because they finally got actual magic and before that the best they could do was feed on the scraps they got from people arguing in a diner. The sentence directly points to them indeed having been around for ages, and the answer to it being "because there's magic now" ties into the movie's theme of Sunset regretting and having to face the consequences of her actions. Go rewatch RR right now.
Out of curiousity, how many people use SillyTavern as opposed to the Visual Novel?
>>43448188From both the general's origins and the number of logs we get, I would say that we have more Silly users than VN ones.But I imagine most anons use both depending on mood or what they want in the moment.
>>43419510Rarity's Little Brother returns - with shimmy shimmy shimmyay shimmyay smimmya ya (Sunset)
>>43448238Log 2
>>43448250Log 3
>>43448257Log 4
>>43448188I haven't touched the VN in months, mostly because I tend to gravitate toward longform writing over short RP interactions. SillyTavern is just better suited to my use case, plus the VN's structure feels a lot more finicky getting the right settings down per model. Better to just avoid the headache entirely.
>>43448188I flitter between the two, though I tend to use ST when I want pone on the go.
Is the new glm worth anything compared to kimi?
>>43444281It is a real shame cause in roleplay and interaction connectivity its better than most DMs. it doesnt need to slowly think and come up with stuff after all. but actually running it, just goes dumb really fucking quick.maybe if you do some crazy shit like running 3 different Fable agents at the same time and burning 1k dollars an hour on OpenRouter just to get a functional game, then it could work. but otherwise you cant get deeper than like 20 messages into ai DMing.
>>43448596positivity bias is unironically the thing that brings 90% of the difficulty with an ai DM. Even if you bring your own rolls for everything it will still fight tooth and nail to make you the unbeatable strongest hero that survives everything and makes every npc bow to you no matter what.
>>43447389Added, thanks Anon.https://mlpchag.neocities.org/view?card=pashatehink/The%20Dazzlings.png
>>43419506How exactly does the lorebook work for the VN? Can I just import one from the Ponydex or does it require more work then that?
GLM5.3 has SO much better writing than 5.2 btw
>>43449185You know what I'm gonna ask for anon.
>>43449185I never really understood what people saw in GLM 5.2. It doesn't really hold a candle to the top stuff, and DeepSeek V4 handles extreme degeneracy much better without the fuss.
>>43449275It's really not bad for the price if you want a model with more smarts and knowledge than old DeepSeek V4. After that, there's quite a jump for stuff like Claude or Kimi.
>>43449066Pretty just import the json file. There's additional options by the basic pony structure works. Kind of.
I'd use the VN more if it had better prompts, i don't feel like messing with it.
>>43449185It's so great that no one even posted any logs of it.
>>43448266Thanks!
>>43449275the new v4 has been a massive step down for me so far. 5.2 still writes better than it most of the time from my testing
>>43449592I like 5.2s writing more but god it is so insanely dumb at logic related tasks
>>43449275are people using really bad sources or something? GLM's man advantage is that its so much smarter than deepseek / kimi.
>>43449561I don't share logs, sorry.
>>43449410Anon... You can just make your own prompts... Hell, I just added elements from Chaggy in, just parts of it I wanted onThe only thing you really have to adhere to is the formatting.Saying that is like saying you don't like subway sandwiches- YOU MADE THE SANDWICH
>>43450188the base prompts are shit
>>43449760What preset do you use for glm?
>>43443769Update, some menus are game breaking because they're unscrollable, the dream selection is what got me just now
>>43450623She's right.
>>43450653From experience?
>>43450641Will look into it.
is linkapi legit?
>>43451030yes ive been using it for months. It goes down sometimes though
>>43451030In what sense?They provide models at a really cheap price, I've been using their opus 4.5 for months it's my go to when the proxies I use are down like right now
>>43451036>In what sense?model availability and whether the models are too scuffed or not what they say they are
>>43451044All the Rarity's little brother logs are of me using opus 4.5 on LinkAPI if you're curious. >>43448238Availability is decent, just dont use the SP mdoels at the topGLM 5.3 is still kinda iffy there from my experience, I did use fable there before they blasted the api costs to infinityJust stick to the constant price models and you're golden, like [次]claude-opus-4-5
>>43451051>>43451044I never actually tested it to see what the model says it is, tried a couple of times telling it to copy my message 1:1 and write it back and from what I see it's clean, but no clue if they just have a workaround for that
>>43451074When does new bread come usually?
>>43451078When the thread is on page 7-8 and at bump limit.
>>43451051>Just stick to the constant price models and you're golden, like [次]claude-opus-4-5Unless you have monstrously large context, you realize you're spending more per-request than you otherwise would on the token-based models right?
New Bread>>43451375>>43451375>>43451375
