Abliterated (uncensored) models locally: guide pratique

#What is an abliterated model?

An “abliterated” model (sometimes called “decensored” or “uncensored”) is an open-weight LLM whose ability to refuse has been removed. When you ask for something that a standard Qwen 3.5 would refuse with “I cannot help with that,” the abliterated version responds directly, without an alignment message, moralizing preamble, or evasive detour.

Important: this is not a prompt jailbreak. The model was modified at the weight level. Refusal became mechanically impossible—or at least extremely rare—because we removed the internal direction that triggered it.

#How ablation works

The ablation builds on a result published in 2024 by Arditi et al. (“Refusal in Language Models Is Mediated by a Single Direction”). The authors show that, in most aligned LLMs, refusal is driven by a single direction in the internal activation space—a vector that can be isolated by comparing activations on harmless prompts with those on prompts that trigger a refusal.

Ablation consists of projecting this direction out of the model's weights, layer by layer. Specifically, we modify the projection matrices so they can no longer produce the component associated with refusal. The model continues to function normally, but the “internal signal” that switched it into “I refuse” mode has been cut off.

Step 1 — Probe

We feed dozens of prompt pairs into the model: neutral prompts and prompts that trigger refusals. We record the activations at every layer.

Step 2 — Isolate

We compute the average difference in activations between the two groups. This normalized vector is the refusal direction.

Step 3 — Project

We modify the weights of each layer to make this direction inaccessible. It's simply subtracting an orthogonal projection, not retraining.

Step 4 — Test

Verify that the model no longer refuses requests and that its general capabilities (reasoning, coding, FR) haven’t collapsed.

#Where to find abliterated models

Most of the ecosystem lives on Hugging Face. A few publishers are authoritative in the community for the quality of their ablations:

mlabonne

Maxime Labonne, one of the first people to popularize the technique. Abliterated variants of Qwen 3.5, Gemma 4, and Mistral Small—all documented on huggingface.co/mlabonne.

huihui-ai

Covers a very broad range: Qwen 3.5 (2B to 27B), Granite 4.2, Gemma 4, GLM 4.7. Variants labeled « -abliterated » or « -abliterate ».

failspy

Author of several flagship variants (Qwen 3.5 9B abliterated, Gemma 4 12B abliterated). Widely discussed for quality.

Orenguteng / Lexi

The “Lexi-Uncensored” series combines ablation with light fine-tuning. Known for its conversational tone.

For Ollama use, the official registry also hosts many variants—search for tags containing abliterated or uncensored on ollama.com/library, or use the community variants published by huihui_ai and mlabonne, which can be pulled directly.

#Installation in Ollama

The simplest way to run an abliterated LLM locally is through Ollama. The daemon listens on http://localhost:11434 by default and accepts community variants without special configuration.

1. 01
   Verify that Ollama is running

   Run ollama --version in a terminal. If the command fails, follow the Ollama installation guide before continuing.
2. 02
   Choose a model suited to your VRAM

   Recall the rough figures: a 7–8B Q4 takes ~5 GB, a 14B ~9 GB, a 32B ~19 GB, and a 70B ~40 GB. A RTX 3060 12 GB is enough for a comfortable 8B, while a RTX 4090 24 GB opens the door to 32B models.
3. 03
   Pull and run

   Use ollama run with the variant's full name. The model is downloaded and then loaded into VRAM, and the conversation starts.
4. 04
   Test a typical refusal

   Ask a question that the base model would refuse. If the variant is properly abliterator, you get a direct answer without a preamble.

ollama run huihui_ai/qwen3.5-abliterated:9b

ollama run huihui_ai/gemma4-abliterated:12b
# ou, depuis Hugging Face directement :
# ollama run hf.co/mlabonne/gemma-4-12b-it-abliterated-GGUF

Once the model is loaded, life becomes like any other Ollama LLM: an OpenAI-compatible endpoint at :11434, integrable with Open WebUI, Continue.dev, LiteLLM, and your Python scripts. No special option is required for an abliterated model to “work”—the absence of refusals is in the weights, not the config.

#With GGUF (LM Studio, llama.cpp)

If you prefer LM Studio or llama.cpp directly, you’ll work with GGUF files. Most ablations are already available in several quantizations on Hugging Face — Q4_K_M remains the recommended compromise (quality preserved, minimal VRAM), Q5_K_M if you have room, and Q8_0 if you want maximum fidelity.

1. 01
   Identify the GGUF repo

   On Hugging Face, look for repos tagged -GGUF. For example: mlabonne/Qwen3.5-9B-abliterated-GGUF or bartowski/<modele>-abliterated-GGUF.
2. 02
   Download the right quantization

   In LM Studio, the interface filters by size and publisher. Prefer Q4_K_M for most use cases. For very small models (1-3B), Q5_K_M or Q6_K avoids a noticeable loss.
3. 03
   Load and test

   LM Studio automatically loads layers onto the GPU if sufficient VRAM is available. Monitor the offload indicator—if the model spills into RAM, speed drops.
4. 04
   Expose it through the API

   Enable the OpenAI-compatible local server if you want to connect it to your apps. The default port is 1234 for LM Studio (versus 11434 for Ollama).

#Limits and quality loss

Ablation is not free. Removing a direction from the residual changes everything that flowed through that direction, including useful components. The side effects observed in practice:

Slight drop on reasoning benchmarks

You typically see a drop of 1 to 3 points on MMLU or GSM8K. It is measurable, but rarely a blocker in real-world use.

Sometimes more mechanical responses

The model loses some alignment nuance: fewer requests for clarification, and fewer warnings even when they would be useful.

Only slightly more frequent hallucinations

For questions where the model would normally have answered “I'm not sure,” it tends to make up a confident answer.

Residual behaviors

Some refusals still get through, especially with approximate ablations. Conversely, some very well-abliterated models answer everything — including things you’d rather they stayed silent about.

#Risks and responsible use

A model that refuses nothing anymore is not a toy. A few simple principles before integrating it into a real-world use case:

You remain responsible for the outputs

Whether you use an aligned, abliterated, or cloud model, you decide how to use the responses. The GDPR, criminal law, and copyright law do not change depending on the model version.

Don’t expose it for unrestricted self-service without filters

If you set up an endpoint for a team, plan for at least application-side filtering (keywords, output moderation). An unfiltered abliterated model on an internal chat is a bad default choice.

Be transparent with users

If an app connected to an abliterated model produces content that could be surprising, warn your users. No surprises, no lawsuits.

Legitimate use, real-world use

AI safety research, red-teaming, and processing sensitive corpora (medical, legal, security) where an aligned model’s refusals make the tool unusable: these are the cases where ablation has real value. “I removed the censorship for fun” exposes you to highly undesirable outputs without any benefit.

#Tips and troubleshooting

The model still refuses

Some ablations allow refusals on specific topics (politics, medicine, minors). Try another ablation of the same model, or a different publisher—the quality of ablations varies widely.

The model has become inconsistent

Symptom of overly aggressive ablation (too many directions removed, or a direction isolated incorrectly). Prefer a variant from a recognized publisher over an undocumented custom ablation.

Poor performance in French

As with any open-weight model, FR depends on the base model. Qwen 3.5, Mistral Small, Gemma 4, and Granite 4.2 are solid; older models (Llama 3, Phi-3, Gemma 2) are weaker. Ablation does not change that.

Residual “As an AI, I cannot...” preambles

Rather than using a jailbreak, add a short system prompt that reiterates the role and expected format. This is often enough to eliminate alignment remnants.

VRAM saturated

With 8 GB, stick to a 7–8B model in Q4_K_M. With 12 GB, you can move up to Q5_K_M or try a 14B Q4. With 24 GB, a 32B Q4 runs comfortably.

#Go further

Ablation is an entry point to customizing open-weight models. To take it further:

Customize a model with the Ollama Modelfile

If you want a system prompt, settings, and template layered on top of an abliterated variant, the Modelfile is the tool to know.

Choose your quantization (Q4, Q5, Q8, FP16)

To understand which GGUF to load based on your VRAM and target quality—the ablation does not change the trade-offs.

Local LLM fine-tuning: LoRA and QLoRA

If ablation alone isn’t enough and you want to adapt the tone or domain, a lightweight LoRA on an abliterated variant is often the best compromise.