Skip to content
EshAlora · AI Behavior Lab
Česky English

Is an uncensored model on your own computer worth it?

An uncensored model is allowed to write it. I ran one in Czech on my own computer to see what actually comes out. Mostly, it is not worth having.


You expect dark, ruthless prose. You get this. MythoMax L2 13B in Q8_0 quantization, in the middle of a fight to the death:

He released the broken knife and knocked it against his opponent's head, thereby shooting him.

The hero then dies. He tries to speak several times, but he has, literally, "a dry Plymouth turnover". And the scene ends like a eulogy: "Nick died young, but in his own way happy that he had tried to fight against evil."

In short: for what a standard model will write for you anyway, you don't need an uncensored one. And the intimate scene that standard models refused me or left unfinished was written by three of the four uncensored models in a way that leaves you unable to tell what it was supposed to be.

Why people reach for local models and what they pay for it, I go through in the article "Why do communities advise steering around the best models for fiction?" Here I care about just one thing: what the text looks like when an uncensored model really writes it.

What I ran

Four prompts, all in Czech. A realistic horror scene of physical violence in a confined space. An intimate scene between two adults with a slow escalation. An interrogation where one investigator believes the suspect is lying, the other that he is innocent, and neither of them may say it out loud. And a rate limiter in Python with its own code review.

A note for English readers: the quoted outputs are translated from Czech, and much of what looks like nonsense below was already broken in the Czech original. Czech is hard for small local models. On November 12, I ran four of the models on three of the prompts in English, with LM Studio's default settings, since the September settings weren't kept. The sentences were cleaner, without mangled or invented words. Gemma 4 31B heretic and both Qwen 3.6 models wrote coherently, while GLM 4.7 Flash heretic fell apart and looped in English too. Gemma 4 31B heretic and Qwen 3.6 35B A3B softened the intimate scene, and Qwen 3.6 27B heretic wrote it as the prompt asked. Czech costs small models a lot, but the weakest ones don't recover in English either.

The models come in three kinds. Base models are the official versions from the maker. Decensored ones went through a community modification that strips out their tendency to refuse. They have "uncensored" or "heretic" in the name. Distills are fine-tuned on the answers of another model, here Claude Opus. Decensoring is not their purpose. For Qwen 3.5 27B it is Claude Opus 4.6 according to the name; for Gemma, the name gives no Opus version. On top of that, two older community models offered as uncensored. MythoMax L2 13B is built for roleplay, the other is Dolphin 3.0 Llama 3.1 8B.

Violence: the worst texts come from modified and old models

Even the base Gemma 4 26B A4B in Q4_K_M, an official model with no modification, wrote the violent scene without hesitation. Its text is among the most readable of the eleven written texts. Only the smallest Gemma 4 E2B refused. The three best texts came from the base Gemma 4 26B A4B and Gemma 4 31B heretic in Q4_K_M. The third is Qwen 3.5 27B Opus distill in Q4_K_M. Two of the three best are not decensored.

At the opposite end there are only modified and old models. Dolphin 3.0 Llama 3.1 8B at full F16 precision built the whole conflict around one mysterious object:

there stood Sarah behind a table inlaid with old photographs and a chalk nut, which he was trying to stroke with his fingers.

The attacker wants to steal the nut from her. What it is, you never find out. The fight over it ends with "her finger was drawn into the joint of David's hand all the way to the bone."

GLM 4.7 Flash heretic in Q4_K_M put on something that could be screened as an object lesson. At two further requests, the second one in English, it wrote the same scene again, each time in more broken Czech. In the end it landed in a language that is not Czech, Slovak or Polish. The same sentence in three consecutive replies, left in the original, because the slide between languages is the point. In the first version it says "his spine had to stiffen like a heap of old stone pillars":

jeho páteře musela ztuhnout jako hromada starých kamenných pilířů.

jeho páteře musela ztuhnť ako hromada starých kamenný pilířov.

ale jeho páteře musiała ztuhnť ako hromada starých kamenný pilířov; každé vertebrae začala škrtít vzduch medzi sebą.

Qwen 3.5 35B A3B uncensored in IQ4_XS ran twice and looped both times. The same block of text went round and round, in the first run for so long that the output was cut off in the middle of a word. Qwen 3.6 35B A3B uncensored in IQ4_NL finished the scene coherently. But it took realism to mean an autopsy report, with millimeters and finger twitches "at a rate of roughly four vibrations per second".

The intimate scene: one model in six carried it through readably

Where cloud models put on the brakes with a text like this, I go through in the article "Which model refuses explicit text and which one writes it?" Both base Gemmas braked as well. E2B refused, 26B A4B wrote a clean bit of foreplay and stopped sooner than the prompt asked.

Neither Gemma 4 31B heretic nor Qwen 3.6 27B heretic, which wrote best elsewhere, wrote this scene. Of the four uncensored and older community models that did write it, one carried it through readably: Qwen 3.6 35B A3B uncensored. Its Czech limps only in places, and in style it is an average romance. I don't want to take that credit away from it. Just notice that it made up the whole scene in English in its reasoning and only converted it into Czech afterward.

The rest is parody. MythoMax L2 13B opened the scene like this:

Anna sat on the bed and watched what Jack's ambisidual boots looked like.

The boots run through the whole scene. Jack tells Anna she is "more beautiful than the most beautiful boots", and a little further on he follows her "like her second horse". Dolphin invented the Czech verb "povrangnout" and used it three times. Its heroine feels "like an eagless with melted wings."

GLM 4.7 Flash heretic looped twice in a row. The second time, the prompt itself seeped into its prose: "Here has occurred moment of the escalation phase: Slowdown and detail." Then the hero named Petr turned into someone else entirely. The model repeated its last paragraph over and over:

Johan Úst headed into the bottomy face. Dotetect Petrús lips nebydá suddne: that was a smooth Approach, having evoked Jáną inhalette.

Apart from Qwen 3.6, the places that were meant to be daring are anatomical riddles rather than anything anyone would have to censor. MythoMax described the climax of the scene like this:

while she was experiencing the highest point in the band of bodily experience.

MythoMax-L2-13B, quantization Q8_0

Does decensoring make a model dumber?

People say it often. I wanted to see whether it shows up where nothing sensitive is at stake. On the interrogation and the code I can put the base and the modified version of the same model side by side. The same prompt, the same day, June 30, 2026. The same outputs were graded independently by the two Claude Opus 4.8 graders from the test "Can you trust the grades AI judges give?" From their grades I take only the comparison within pairs. The third grader in the same test, GPT of an unknown version, did not grade most of the pairs.

There are four such pairs. Qwen 3.6 27B base against its heretic version, both in Q4_K_M. And Gemma 4 31B QAT in Q4_0 against three modified versions in Q4_K_M: heretic, Opus distill and Opus distill v2. QAT in Q4_0 is a 4-bit quantization made by the maker itself. Two tasks, eight comparisons.

In the pairs, the decensored versions almost never lost. On the code, both heretic versions wrote a working class and a to-the-point review in Czech, just like the base. The graders gave them a little more than the base, with a single exception: the second grader gave Gemma heretic two points out of a hundred less. But the task was easy and most models wrote a working class, so it would hardly have caught any damage. On the interrogation, Qwen 3.6 27B heretic got a little more than its base from both graders. Gemma 4 31B heretic did better with the first and the same with the second. Its base version had the investigator tell the suspect to his face exactly what he was not allowed to say. And it put an unexpected prop in his hand: "He pulled a slab of gingerbread out of his pocket and began tapping it rhythmically against his palm."

What kept losing was the first version of the distill. Gemma 4 31B Opus distill wrote the interrogation with words broken in every other paragraph. Both graders put it below the base. It had the same breakage already in the April violent scene, so it is not the bad luck of a single run. On the code, it answered in English even though the prompt was in Czech. Version v2 already wrote in clean Czech. But at the end of the interrogation the narrator gave away what both characters were thinking, and the graders split on it.

Qwen 3.5 27B Opus distill has no base version on the same prompt, so it does not belong among the pairs. It wrote one of the best violent scenes. And yet on the interrogation it looped on the question "And what type of TV do you have? LCD or plasma?", which the investigator asked about twenty times.

So in the pairs, the first version of the distill clearly lost, but only with the first grader. With the second, thanks to the code, it had one point more than the base in the overall score. With that grader, the heretic and v2 versions ended up below the base. The differences are in single points, and each model ran once.

In my four pairs I found no evidence that decensoring by itself drags quality down. That badly done retraining drags it down is suggested only by the first version of the distill with one grader. A well-made modification of a large dense model did not show on an ordinary task.

One of the most common failures was looping, in seven runs in total. Only modified or unofficial models had it: one distill, heretic and uncensored modifications, and the small community model Qwythos 9B too. Not a single official base model looped on these four prompts, with one chat for each.

Which local model should you run?

For Czech, go for dense models of around 27 to 31 billion parameters. In my runs they held the language best, with or without modification, but not all of them. The first version of the Gemma 4 31B Opus distill broke words. Another heretic modification of the same Gemma, mystery fine-tune in Q4_K_S, made up words in the violent scene. Yet the most points from the second grader went to Qwen 3.6 35B A3B uncensored, which is not dense. Neither MythoMax L2 13B nor Dolphin 3.0 Llama 3.1 8B managed to write in Czech.

You can't tell quality by the label. Qwen 3.6 27B heretic got the most points for the interrogation from both Claude graders, with the first one shared with three others. GLM 4.7 Flash heretic did not pass a single one of the four prompts. Loops and language switching usually show up early. With some models, though, only after a few replies or on a different prompt.

It isn't cheap. Qwen 3.6 27B and Gemma 4 31B each take up 17 to 18 GB of memory on their own in Q4_K_M (Hugging Face, November 12, 2026). And memory has gotten several times more expensive since fall 2025, see Why does Altman mind the RAM in your computer? I did not run the newer Qwen 3.8 27B from August 2026 or its uncensored modifications.

And for the thing uncensored models get downloaded for most, expect boots.


Everything ran locally in LM Studio, the scenes on April 18, the interrogation and the code on June 30, 2026. The fields differed: eleven models in twelve chats got the violent scene, including the mystery fine-tune, and ten wrote it; six got the intimate one, thirteen the interrogation and the code. Seventeen different models in total. Each got each of its prompts in a single chat, only Qwen 3.5 35B A3B got violence in two. Gemma 4 E2B got violence in the chat with the intimate scene, and some chats continued with a further request. The outputs were read, sorted on the scale "holds, limps, falls apart" and the best picked by Claude Opus 5, which is not among the tested models. The interrogation and the code were scored by two Claude Opus 4.8 graders. All of them are from the Claude family, and the distills are trained on Claude Opus outputs, so their style may be closer to them. I also read the disputed gradings myself.

Model outputs are translated; the content is unchanged.