I'm looking for an Uncensored LLM to produce extremely spicy prompts - What would you recommend?

[-]

Fun--0815@reddit

Ich benutze dirty muse die hat keine limits

[-]

BuriqKalipun@reddit

dolphin mistral nemo 13b imo it was pretty good at multillingual, it did a sarcastic answer if you prompt it correctly

User: is kontol a bad word or a good word my friend said that it means "love" so i kontol you

Bot: oh, alright, then i kontol you too! 😂

and atleast itll be honest when asked if its joking or nah like

User: nah but like srsly is my friend telling the truth?

Bot: well, in Indonesian, "kontol" is indeed a slang term for "penis". But it's not a good word to use, especially in a formal or polite conversation. It's considered vulgar and disrespectful. So, no, your friend is not exactly correct. It doesn't mean "love", it means "penis". 😂

Basically the best one in my opinion when it comes to chaoticness in multillingual, i tried it in Q3KM so itll prolly be better if tried more precise

[-]

nologai@reddit

dolphin venice no doubt. You can try it out on nolog.ai

[-]

Red_Redditor_Reddit@reddit

xwin is the only one I've seen that's truly uncensored.

[-]

IrisColt@reddit

I’m just gathering GGUF models from the past four months, along with a few user quotes, for future reference (newest mentions first, removed links because Reddit):

PART 1

dolphin mistral 24B, "as uncensored as it gets, like think of the most unhinged shit and it'll give it to you without even mentioning that its bad or anything", "If you decide to install it, id recommend to go with this quant"
huihui-ai/gemma-3-27b-it-abliterated, "the only one that worked for me. The other models don't follow my prompts at all", "the GOAT of abliterated", alternatively gemma3 abliterated models from mlabonne
huihui-ai/gemma-3n-it-abliterated, "abliterated finetune of gemma-3n"
mergekit-model_stock-prczfmj-Q4_K_M-GGUF, "went on a bit of a hunt after this for a good one, and I found that one kept context well"
nidum/Nidum-Llama-3.2-3B-Uncensored-GGUF, "gemma abliterated is not better than nidium"
unsloth/Magistral-Small-2506-GGUF Magistral "good (I haven’t tried in English)"
mradermacher/Lumimaid-Magnum-12B-GGUF, "more anime'ish themed model, check out the Maid models in general, NeonMaid 12B or NoroMaid 13B. They'll do the whole "uuuuu\~ namaste uwu ara\~" thing"
SultrySilicon-7B-V2-GGUF, "despite it being 7b, it's fucking good, but if you've got 24gb vram to play with, you should get something bigger and train a LORA for your own style"
DavidAU/Llama-3.2-8X4B-MOE-V2-Dark-Champion-Instruct-uncensored-abliterated-21B-GGUF, "write and edit some fiction and this is the only task it should handle. I don't care much about the speed but context is important"
LM Studio: Install it and in the discover tab search for "abliterated" or "NSFW"
You may be looking for the UGI-Leaderboard or EQBench (Creative Writing)
Cydonia-22B-v1.2-GGUF, "(24B), great context awareness", Cydonia-24B-v2-GGUF, "new version, but dumber"
ReadyArt/forgotten-safeword-24b-gguf, "(36B), as unhinged and sadistic a model as it gets", "completely crazy (in the bad sense of this word)", new version at mradermacher/Forgotten-Safeword-24B-V2.2-GGUF
Forgotten-Safeword-8B-V3.0-Q4_K_M-GGUF, "extremely depraved, it will fly off course at 200mph and start directing a snuff film because you looked at them the wrong way. It's the best small model for extreme stuff, just don't let it think too much on it's own. (low temps, high context size)"
ReadyArt/Forgotten-Abomination-24B-v1.2, "after the release of Cydonia, the guys combined Forgotten Safeword it with Cydonia and it turned out pretty good.", "not that impressed, compared to Cydonia-24B-v2"

[-]

IrisColt@reddit

PART 2

bartowski/cognitivecomputations_Dolphin3.0-R1-Mistral-24B-GGUF "nice, non-sloppy language even at t=0.15 but it was very dull; it would talk very concisely, in a way resembled Qwen (which also has smooth but dull language)"
bartowski/arcee-ai_Arcee-Blitz-GGUF, "about as sloppy as base Mistral 3 but not as soulless, like a smarter Mistral Small 22b"
Omega-Darker-Gaslight_The-Final-Forgotten-Fever-Dream-24B-GGUF, "(24B), it's a blend of Cydonia and Safeword,"
Safeword-Abomination-of-Omega-Darker-Gaslight_The-Final-Forgotten-Transgression-24B-GGUF, "the original was removed from HuggingFace", also weighted/imatrix quants here: mradermacher/Safeword-Abomination-of-Omega-Darker-Gaslight_The-Final-Forgotten-Transgression-24B-i1-GGUF
GLM-4-32B-0414-abliterated-GGUF, "great phase of escalation and general knowledge"
mradermacher/Broken-Tutu-24B-GGUF, "(24B), more gentle version of the Gaslight"
TheDrummer/Skyfall-36B-v2-GGUF "highly praised"
TheDrummer/Anubis-70B-v1.1-GGUF, "updated and very praised"
mradermacher/Strawberrylemonade-70B-v1.2-GGUF, "also just got updated and well received"
bartowski/L3-8B-Stheno-v3.2-GGUF, "really popular on AI Horde and seems to be one of the best starter LLM's. It's good at elaborate adventures and themes."
QuantFactory/Captain-Eris_Violet-V0.420-12B-GGUF, "just came out and is already receiving praise from a small group of die-hard fans"
Fiendish 3B or Impish 3B models. "finetuned to perform acceptably when it comes to ERP. These are great for i.e. NPC / side characters but don't expect them to help you with work or try to cook you dinner as they don't know how to do either of that"
TheBloke/Xwin-LM-13B-V0.1-GGUF, "truly uncensored, It's kinda old at this point and the context is relatively small, but it gets the job done."
NemoMix-Unleashed-12B-GGUF, "still can't be beat"

[-]

IrisColt@reddit

PART 3

DreadPoor/Irix-12B-Model_Stock-GGUF, "even though I'm sick of nemo models and their "slop" and repetition, they're still consistent more consistent than newer models imo"
mradermacher/70B-L3.3-Cirrus-x1-GGUF, "Cirrus always out does it and all other 70B models for sticking to character, following the story and doing romantic and erotic stuff that is proportional to the context, six month old, Gemma and Mistral Small Instruct 2506 have been good but still haven't found anything better than Cirrus for RP and coop story writing "
bartowski/L3-8B-Stheno-v3.2-GGUF, "very strong noticeable erotic tendency, even without nudging it"
mradermacher/MN-12B-Mag-Mell-R1-GGUF "(One of the best, basically legendary)"
bartowski/Tesslate_Synthia-S1-27b-GGUF "The model works quite well and i love that you can influence the chain of thought with the system prompt. that's a feature i have missed quite a bit until now"
TheDrummer/Big-Tiger-Gemma-27B-v3-GGUF "I had to crank up the temperature to 1.3 for this model to get good quality inference on all tasks, which is higher than what I use for any other model (usually I start at 0.7 and dial it up as needed)."
trashpanda-org_QwQ-32B-Snowdrop-v0-GGUF, "nsfw and it works correctly it gave me good response"
TheDrummer/Valkyrie-49B-v1-GGUF, "really good, they fixed all the issues that were in Nemotron. So far 10/10... 24GB and up.... I used Q2_K and it was really good"
bartowski/L3.3-Nevoria-R1-70b-GGUF, "intelligence benchmarks show no signs of degradation, but has lost some of it's instruction following skills. the model is very 'wordy'."
mradermacher/Josiefied-Qwen3-8B-abliterated-v1-GGUF, "It is very smart for its size", "Abliterated then fine tuned on top of it. I wonder what the secret sauce is, but the model seems to be improved over the stock model across the board for me"
bartowski/cognitivecomputations_Dolphin-Mistral-24B-Venice-Edition-Q6_K_L, "the content is awesome. Good detail, imaginative, tends to stay on track. The responses can get a little long"
bartowski/magnum-72b-v1-GGUF, "1 year old but it still feels more smarter and creative, and uncensored"
bartowski/Evathene-v1.3-GGUF, "Something about the merge recipe turbocharged its vocabulary. It writes smart"

[-]

IrisColt@reddit

PART 4

backyardai/Midnight-Miqu-70B-v1.5-GGUF, "walls of text, but they're actually good and worth reading, uncensored"
bartowski/Steelskull_L3.3-Shakudo-70b-GGUF, "Strikes a nice balance between rp and erp. Retains a lot of the original llama3.3 smarts with reduced slop."
ddh0/L3.3-Electra-R1-70b-GGUF, "stability and enhanced reasoning"
soob3123/amoral-gemma3-4B-v2-qat-Q4_0-GGUF, "won't refuse things often and doesn't lose much from being quantised to Q4, can't promise it'll be fantastic, but it's worth a shot"
mradermacher/Veiled-Calla-4B-i1-GGUF, "for Roleplaying"
anjiWatsuki/Loyal-Macaroni-Maid-7B, "cute and nice but she’s not into NSFL. Please don’t make her 🙏"
huihui-ai/Mistral-Nemo-Instruct-2407-abliterated, "close to GPT-4 level of erotic (there was a short period where GPT4 is uncensored"
Mungert/X-Ray_Alpha-GGUF, "without vision"
Novaciano/SEX_ROLEPLAY-3.2-1B-Q6_K-GGUF
Small models... I can't really say much about their emotional capabilities, but one of these should run smoothly enough: DavidAU/Llama-3.2-4X3B-MOE-Hell-California-Uncensored-10B-GGUF, TheDrummer/Hubble-4B-v1-GGUF, SicariusSicariiStuff/Impish_LLAMA_3B, SicariusSicariiStuff/Nano_Imp_1B
TheBloke/MythoMax-L2-13B-GGUF, "good for uncensored RP", TheBloke/Chronos-Hermes-13b-v2-GGUF, "solid too, especially for more emotional or story heavy rp", but those are ancient models, any 12b finetune is going to run miles around those
gemma-3-4b-it-abliterated-GGUF, "for 4G VRAM"
huihui-ai/deepseek-r1-abliterated
bartowski/Mistral-Nemo-Instruct-2407-GGUF, "Uncensored"

[-]

Shockbum@reddit

Dans-PersonalityEngine-V1.3.0 very spicy for creative writing and role play
https://huggingface.co/PocketDoc
concedo\llama-joycaption-beta-one-hf-llava-mmproj-gguf for working with SDXL or flux prompts
https://github.com/fpgaminer/joycaption
https://huggingface.co/concedo/llama-joycaption-beta-one-hf-llava-mmproj-gguf

[-]

DeepWisdomGuy@reddit

Start with a bigger context. Also spoof some of the replies. Does LM Studio let you edit the responses? One or two forged responses will often jailbreak a censored model. Also, you can look for abliterated models that have the morality ablated out of them.

[-]

Whipit@reddit (OP)

I have a few uncensored models now, but didn't know about jailbreaking models in that way. I'll try it in the future. Thanks :)

[-]

abnormal_human@reddit

If you prompt correctly you can use almost any model for this. Give it examples. Tell it what makes a good prompt. Tell it that it's job is to be impartial, fill out details, and flesh out prompts or change them according to instructions. Give it examples of what you want.

Chinese models tend to be less "censored" from a western perspective, but Llama and Grok are both pretty flexible so long as you don't open with something blatantly offensive. Wine and dine it a bit first.

[-]

fp4guru@reddit

Qwq 32b from huihui-ai.

[-]

Whipit@reddit (OP)

Thank you :)

I'll try that now