nsfwapis.com
// Guide / Models

What is an uncensored LLM?

An uncensored LLM is a language model that has been changed so it no longer refuses requests its original version was trained to decline. There are three common ways to make one: retrain it on data without refusals, edit the refusal out of its weights, or merge it with other models. This guide explains each method, the model families you will see most, and when running one locally beats calling it through an API.

What is an uncensored LLM?

An uncensored LLM is a large language model whose refusal behavior has been reduced or removed after its original release. You ask it something the base chat model would decline, such as adult fiction or blunt answers on a sensitive topic, and it answers instead of replying “I can’t help with that.”

Almost every uncensored model starts from an open-weight release. Labs like Meta, Alibaba (Qwen), Zhipu (GLM), DeepSeek, Moonshot (Kimi), Mistral and Google (Gemma) publish their weights,, so anyone can change them. Closed models can’t be uncensored this way, because nobody outside the lab has the weights.

“Uncensored” isn’t a technical term. Model builders and API providers use it for several different things:

  • A model fine-tuned on data that contains no refusals.
  • An abliterated model, with its refusal direction removed from the weights.
  • A merge of several models, one or more of them uncensored.
  • A standard model served with a permissive system prompt.
  • An API that skips its own moderation filter, whatever model sits behind it.

Only the first three change the model itself. When you see the word on a model card or provider listing, look for the base model and the method.

Why do standard LLMs refuse requests?

Standard LLMs refuse because they were trained to. Refusal is a learned behavior, added after pretraining, not a filter bolted on top.

A base model is first pretrained on a huge amount of text to predict the next token. At that stage it has no idea of what to refuse. The lab then trains it to follow instructions, usually in two steps: supervised fine-tuning on example conversations, then preference tuning (RLHF or a method like DPO), where the model learns to prefer the answers people rated better. Both steps include examples where the right answer is a refusal, a warning or a lecture.

That has two consequences that matter for uncensoring:

  • The refusal behavior lives in the weights. You can change it by changing the weights, which is what every real uncensoring method does.
  • Refusals spill over. A model trained to refuse some requests often refuses harmless ones that look similar, such as a crime novel or a medical question about drugs. Fixing these false refusals is a big part of why people want uncensored models at all.

Hosted products often add a separate moderation model that checks prompts and replies. Uncensoring the LLM does nothing to that layer.

How uncensored fine-tunes are made

An uncensored fine-tune is an existing model trained further on a dataset that contains no refusals. This was the first popular method. In 2023 Eric Hartford published an essay called “Uncensored Models” and a set of WizardLM models trained on an instruction dataset with the refusing and moralizing answers filtered out. The idea was simple: if the model never sees a refusal during instruction tuning, it doesn’t learn to refuse.

The steps are much the same today:

  1. Pick a starting point. A base (pretrained only) model or an instruction-tuned chat model.
  2. Build or filter the dataset. Remove refusals and disclaimers from an instruction dataset, and often add writing the builder wants more of, such as roleplay or long-form fiction.
  3. Train. Most fine-tunes use LoRA or QLoRA, which train a small set of extra weights instead of the whole model. That lets a single GPU fine-tune models that would otherwise need a cluster.
  4. Optionally run preference tuning. Some builders add a DPO pass that rewards direct answers over refusals.

Fine-tunes also change style, length and personality, which is why fiction and roleplay communities often prefer a good fine-tune over a plain abliterated build. The flip side: a narrow or low-quality dataset can make the model repetitive or weaker at general tasks.

Abliteration and model merges

Abliteration and merging are the other two common routes, and both need far less compute than a fine-tune.

Abliteration

Abliteration removes refusals by editing the weights directly, with no training. It comes from 2024 research by Arditi and colleagues showing that refusal in many chat models runs through a single direction in the model’s internal activations. Builders find that direction by comparing activations on prompts the model refuses and prompts it answers, then remove it from every weight matrix that writes into the residual stream. The result is a normal model file with the same size and architecture. The guide to abliterated models covers the method and its quality cost in detail.

Model merges

A merge combines the weights of two or more models built on the same base into one new model, again without training. The open-source tool most builders use is MergeKit. Common methods include:

  • Linear and SLERP. Average the weights of two models, either as a straight line or along a curve that keeps their scale. Simple and widely used.
  • Task arithmetic and TIES. Work out how each fine-tune differs from the base model, then add those differences together. TIES-Merging trims small changes and resolves sign conflicts so the fine-tunes interfere less.
  • DARE. Randomly drops most of each fine-tune’s changes and rescales the rest before merging, which often reduces interference further.

Builders merge an uncensored model with one that writes or reasons better, hoping to keep both traits. Results vary, model cards are often thin, and a merge can bring back refusals from one of its parents.

How the main ways to uncensor an LLM compare
MethodTraining neededWhat it changesMain risk
Uncensored fine-tuneYes, usually LoRA or QLoRARefusals, plus style and tone from the datasetWeaker or repetitive if the dataset is narrow
AbliterationNoRemoves the refusal direction onlySmall drop in reasoning and instruction following
Abliteration + fine-tuneShort DPO or SFT passRefusals removed, then quality recoveredMore work, so fewer builds exist
MergeNoMixes traits of several modelsUnpredictable, sometimes poorly documented

Popular uncensored AI models and families

Uncensored models follow open-weight releases, usually appearing on Hugging Face within days or weeks of a new model. The families you will see most are:

  • Llama, Qwen, Mistral and Gemma. Released in sizes that run on consumer hardware, so they dominate local uncensored models. Most early uncensored fine-tunes were built on Llama.
  • GLM, DeepSeek and Kimi. Large mixture-of-experts models, and the main bases for the hosted uncensored LLMs in this directory.

Here are the uncensored and abliterated models the directory currently lists, grouped by family, with published prices per 1 million tokens:

Hosted uncensored LLMs by model family · USD per 1M tokens
FamilyModelTypeProviderInputOutput
GLMGLM 5.3 Flash AbliteratedAbliteratedRefuseless$0.12$0.50
GLMGLM 5.3 AbliteratedAbliteratedRefuseless$1.00$2.00
GLMGLM-5.3 Flash UncensoredUncensored buildUnaligned$0.063$0.10
GLMGLM-5.3 UncensoredUncensored buildUnaligned$0.15$0.24
GLMGLM-5.3 AbliteratedAbliteratedUnaligned$2.00$3.20
GLMGLM-5.3 Flash Cyber UncensoredUncensored buildUncensored.com$0.96$1.92
GLMGLM-5.3 Flash UncensoredUncensored buildUncensoredLLM$0.1053$0.3685
GLMGLM-5.3 UncensoredUncensored buildUncensoredLLM$0.282$0.846
GLMGLM-5.3 AbliteratedAbliteratedUncensoredLLM$2.1825$4.365
GLMGLM 5.3 Flash UncensoredUncensored buildNanoGPT$0.20$0.80
GLMGLM 5.3 UncensoredUncensored buildNanoGPT$1.25$2.25
DeepSeekDeepSeek V4.1 Flash 1017 AbliteratedAbliteratedRefuseless$0.12$0.45
DeepSeekDeepSeek V4.1 Flash Cyber UncensoredUncensored buildUncensored.com$0.96$1.92
DeepSeekDeepSeek V4 Pro UncensoredUncensored buildUncensoredLLM$0.97$1.94
KimiKimi K3.1 Pro Cyber UncensoredUncensored buildUncensored.com$4.80$24.00
KimiKimi K3.1 Pro UncensoredUncensored buildUncensoredLLM$1.20$6.00

GLM builds are hosted by Refuseless, Unaligned, Uncensored.com, UncensoredLLM and NanoGPT. DeepSeek builds are hosted by Refuseless, Uncensored.com and UncensoredLLM. Kimi builds are hosted by Uncensored.com and UncensoredLLM. Matching names across providers don’t mean matching models. Each provider makes or picks its own build, so a “GLM-5.3 Flash Uncensored” at one provider can behave quite differently from one with a similar name at another. The provider comparison covers prices, payment and compatibility for each.

Can you run an uncensored LLM locally?

Yes, if the model fits in your GPU memory, or in system RAM if you accept slower output. Most uncensored models on Hugging Face are published as GGUF files for llama.cpp and tools built on it, such as Ollama and LM Studio. These files are quantized: each weight is stored in fewer bits, which cuts memory use at some cost in quality.

You can estimate how much memory the weights need with one line of arithmetic:

memory in GB ≈ parameters in billions × bits per weight ÷ 8

Approximate memory for model weights alone, before context and overhead
Model size16-bit8-bit4-bitExample
8B16 GB8 GB4 GBSmall chat models
32B64 GB32 GB16 GBMid-size models
70B140 GB70 GB35 GBLarge dense models

Add a few gigabytes on top for the context (the KV cache grows with every token in the conversation) and for the runtime itself. In practice, a 4-bit 8B model runs on most recent gaming GPUs, a 32B model needs a 24 GB card, and a 70B model needs two large cards or a machine with a lot of unified memory.

The big families behind most hosted uncensored LLMs are another matter. DeepSeek V3, for example, has 671 billion total parameters according to its technical report. Even at 4-bit, a model that size needs hundreds of gigabytes of memory, which is far beyond a home setup. That is the main reason APIs exist for these models.

With Ollama, running a local model is one command (ollama run followed by the model name), and Ollama serves an OpenAI-compatible API on localhost:11434/v1. Code you write against a local model can later point at a hosted one by changing the base URL, key and model ID.

Uncensored LLM API or local model: which is cheaper?

A local model is cheaper only if you already own capable hardware or have a lot of steady traffic. For most people starting out, a hosted API costs less than buying a GPU.

Here is the scale on published rates. Unaligned lists GLM-5.3 Flash Uncensored at $0.063 per 1M input tokens and $0.10 per 1M output tokens. A $10 deposit buys about 159 million input tokens at that rate. A roleplay app sending 50,000 messages a month, at 6,000 input and 300 output tokens each, costs about $20.40 a month on it before cached input.

Running an uncensored LLM locally vs calling it through an API
Local modelHosted uncensored LLM API
Upfront costGPU or machine, unless you own oneNone, prepaid balance from a small deposit
Cost per tokenElectricity onlyPublished per-token rate
Largest model you can runLimited by your memoryLarge models such as GLM, DeepSeek and Kimi builds
PrivacyPrompts never leave your machineDepends on the provider’s logging policy
RulesThe model license and the lawThe provider’s terms, the model license and the law

A practical split: prototype on a small local model, then move to an API when you need a larger model or more users. Both speak the OpenAI format, so switching is a small change. The pricing guide shows how to price your own workload across providers.

What an uncensored LLM doesn’t change

  • Knowledge. Uncensoring removes refusals. It doesn’t add facts or skills the base model lacked.
  • Provider terms. A hosted API sets its own acceptable-use rules and logging policy, whatever model it serves.
  • The license. A fine-tune, abliterated build or merge normally carries the base model’s license. Check it before you ship weights or build a commercial product.
  • The law. Illegal content stays illegal. Sexual content involving minors, and intimate images of real people made without their consent, are illegal and never allowed, whatever model produced them.

How to choose an uncensored model

  1. Find the base model and method. A model card should name the base release and say whether it was fine-tuned, abliterated or merged. Skip models that say neither.
  2. Match the size to where it runs. Use the memory formula above for local models. For anything larger, look at hosted options.
  3. Build a test set. Write 20 to 50 real prompts from your app, including ones a standard model refuses and plenty of ordinary ones.
  4. Score refusals and quality separately. A model that never refuses but writes flat, repetitive text isn’t a win.
  5. Price it. For hosted models, price your real token counts, not the headline input rate.

Frequently asked questions

What is an uncensored LLM?

An uncensored LLM is an open-weight language model that has been modified to refuse far fewer requests than the original. The usual methods are fine-tuning on data without refusals, abliteration, and merging with other models.

Are uncensored LLMs legal?

Downloading and running an open-weight model, including an uncensored one, is generally allowed under its license. What you generate is still subject to the law and to any provider’s terms. Sexual content involving minors and non-consensual intimate images of real people are illegal everywhere.

What is the difference between an uncensored and an abliterated model?

Abliteration is one specific way to uncensor a model, by editing a refusal direction out of its weights without training. “Uncensored” is a broader label that also covers fine-tunes trained on data without refusals and merges.

Can I run an NSFW LLM on my own computer?

Yes, if it fits in memory. A 4-bit model needs roughly half its parameter count in gigabytes, so an 8B model needs about 4 GB plus room for context. Tools like Ollama and LM Studio can run GGUF builds of uncensored models.

Where can I use uncensored AI models through an API?

The directory lists 5 uncensored LLM API providers: Refuseless, Unaligned, Uncensored.com, UncensoredLLM and NanoGPT. Most host uncensored or abliterated builds of GLM, DeepSeek or Kimi and accept requests in the OpenAI format.

Sources

Prices and offers come from each provider’s published information, checked . They can change, so confirm them on the provider’s site. Found something out of date? Tell us.