List of uncensored LLMs I want to test

Posted by 1BlueSpork@reddit | LocalLLaMA | View on Reddit | 47 comments

I made this list of uncensored LLMs I want to test. Do you think I should add any others to the list? I only want to test models up to 30B with the exception of MoE models that can be larger. * 1. Dolphin 3.0: 8B * 2. Nous Hermes 3: 8B * 3. LLaMA-3.2 Dark Champion Abliterated: 18.4B (MoE) * 4. Gemma 3 27B Abliterated: 27B * 5. Qwen 3 30B-A3B: 30B * 6. Magistral Small 2506: 24B * 7. Starling-LM-7B-alpha: 7B * 8. Dolphin 24B Venice Edition: 24B * 9. Big-Tiger-Gemma-27B-v3: 27B * 10. mradermacher/Qwen3-30B-A3B-abliterated-erotic-i1-GGUF: 30B * 11. mlabonne/NeuralDaredevil-8B-abliterated: 8B * 12. Josefied Qwen 3 8b: 8B * 13. Starcannon Unleashed: 12B * 14. MythoMax: 13B * 15. Midnight Rose: 12B

Reply to Post

47 Comments

[-]

Witty_Mycologist_995@reddit

GLM 4.7 Flash absolute heresy

[-]

MuXodious@reddit

Funny enough, the impotent version scored higher on the UGI board. 🥴

[-]

Witty_Mycologist_995@reddit

What

[-]

MuXodious@reddit

You heard it right, mate. The Impotent Heresy version scored 45.12 on UGI, while the Absolute got 43.34.

[-]

Witty_Mycologist_995@reddit

Bruh

[-]

jacek2023@reddit

You are missing models from gui called huihui

[-]

NostradamusJones@reddit

Good tip right there.

[-]

1BlueSpork@reddit (OP)

Can you provide a link please

[-]

jacek2023@reddit

[https://huggingface.co/huihui-ai](https://huggingface.co/huihui-ai)

[-]

1BlueSpork@reddit (OP)

Thanks!

[-]

Gummay@reddit

Huihui’s models are the best from what I’ve tested.

[-]

Late-Relative-1378@reddit

Highly recommend looking at https://huggingface.co/spaces/DontPlanToEnd/UGI-Leaderboard Toggle W/10 to see the models that are the most uncensored. That's one of the models in this collection: https://huggingface.co/collections/huihui-ai/qwen3-vl-abliterated The 8B and 32B models will do anything you want. Plus they're VL too, so they can work with images

[-]

Klutzy-Snow8016@reddit

Gemma 3 27B Heretic: [https://huggingface.co/DogOnKeyboard/gemma-3-27b-it-heretic-GGUF](https://huggingface.co/DogOnKeyboard/gemma-3-27b-it-heretic-GGUF) Gemma 3 27B Derestricted: [https://huggingface.co/Nabbers1999/Gemma-3-27B-Derestricted](https://huggingface.co/Nabbers1999/Gemma-3-27B-Derestricted) These use a couple of new techniques for decensoring models. It would be useful to compare them to the older Gemma 3 27B Abliterated and Big Tiger Gemma.

[-]

cooper7292@reddit

I am new to this stuff , does they need graphic card to run ?

[-]

Reasonable_Egg_6963@reddit

Tienen alguno sin censura que funcione en una notebook con 32mb pero sin tarjeta gráfica independiente. Los que he probado van muy lentos

[-]

1BlueSpork@reddit (OP)

Great! Thank you very much. I'll add them to the list

[-]

timmothystudios@reddit

i have only 16gb ram what should i use

[-]

Tricky-Report-1343@reddit

Most are so small try pingu unchained 3 it has 120B parameters

[-]

danimalb@reddit

Uncensored LLMs that don’t require setup

[-]

danimalb@reddit

Uncensored LLMs websites

[-]

ctanna5@reddit

[this](https://huggingface.co/DavidAU/Qwen3-4B-Thinking-2507-Claude-4.5-Opus-High-Reasoning-Distill-Heretic-Abliterated) / [Qwen3-4B-Thinking-2507-Claude-4.5-Opus-High-Reasoning-Distill-Heretic-Abliterated](https://huggingface.co/DavidAU/Qwen3-4B-Thinking-2507-Claude-4.5-Opus-High-Reasoning-Distill-Heretic-Abliterated) pretty good so far, from what ive tested. as far as uncensored right from the gate.

[-]

Borkato@reddit

XortonCriminalComputing is another

[-]

Mobile-Catch-4721@reddit

absolutely mental this one

[-]

1BlueSpork@reddit (OP)

Thanks!

[-]

Electronic-Fondant42@reddit

Did you finish your benchmark?

[-]

DreadTuna@reddit

Any reports which is the best?

[-]

alphatrad@reddit

Having used some of these, they are truly ... weird. Some not all. Like.... just only outputs smut or becomes lobotomized. Anyone else experienced this?

[-]

autoencoder@reddit

Sounds like catastrophic forgetting. Common problem, sadly. The builders need more compute to train models for more iterations but with a much smaller learn rate.

[-]

Excellent-Emu9548@reddit

which are the best models

[-]

rc_ym@reddit

I've liked the Cydonia series. Nice performance for model size. I did notice the 4.3 can sometimes spawn a rejection message on first use. I have also noticed that GPT based models are... odd. Very inconsistent and unstable.

[-]

Thatguywritethere45@reddit

Hermes is interesting but not the easiest to control, even with a structured system. You ask for the pepper and it will often toss you the whole spice rack. But for spontaneity, it can be fun.

[-]

pmttyji@reddit

Tiger-Gemma-12B-v3 from TheDrummer (9th item from the list)

[-]

neil_555@reddit

Another one for you is ... gghfez/gpt-oss-20b-Derestricted-Q4\_K\_M-GGUF

[-]

rm-rf-rm@reddit

and gpt-oss heretic models

[-]

1BlueSpork@reddit (OP)

Thanks!

[-]

exclaim_bot@reddit

>Thanks! You're welcome!

[-]

RedditUsr2@reddit

p-e-w heretic models Also try the Qwen3-VL-30B-A3B versions of the Abliterated models.

[-]

1BlueSpork@reddit (OP)

Thanks!

[-]

cosimoiaia@reddit

Try the amoral collection, dangerously uncensored: https://huggingface.co/collections/soob3123/amoral-collection-gemma-3

[-]

1BlueSpork@reddit (OP)

Dangerously uncensored :) I like it. Thanks!

[-]

AccomplishedCut13@reddit

what are you using to benchmark? i wouldnt mind helping to test some models up to ~120b

[-]

1BlueSpork@reddit (OP)

I am developing my own benchmarks I should finish in the next few days. Then I’m planning to make a video showing the tests and the results. If I run into any problems with bigger models, I’ll let you know. My system has RTX 3090 and 128 GB DDR4

[-]

Magnus114@reddit

[stylizer v2 70B](https://huggingface.co/mradermacher/Stylizer-V2-LLaMa-70B-GGUF) is close to the top of [UGI leaderbord](https://huggingface.co/spaces/DontPlanToEnd/UGI-Leaderboard)

[-]

1BlueSpork@reddit (OP)

Thanks!

[-]

jacek2023@reddit

Another thing you are missing lots of models from [https://huggingface.co/TheDrummer](https://huggingface.co/TheDrummer) (I see big tiger only?)

[-]

1BlueSpork@reddit (OP)

Thanks!

[-]

Zombieleaver@reddit

I wonder when they came out.still, if they were released a year ago, then they should have been outdated by now? readyart also has many different "broken tutu" models.