# FAQ

> Common questions about abliterated and uncensored models, hardware and local setup.

## What is the difference between “abliterated” and “uncensored”?

**Abliterated** usually refers to a specific weight-editing technique that removes (or greatly weakens) the model’s “refusal direction” without full retraining. The method was popularized in 2024 and is now commonly applied to new open-weight releases within hours or days.

**Uncensored** is a broader, informal label. It can mean:
- An abliterated model
- A model fine-tuned on data that reduces refusals
- A model trained with weaker safety alignment from the start

In practice, many people use the terms interchangeably. Always check the model card for the actual method used (abliteration, Heretic, fine-tune, etc.).

_Tags: basics, terminology · Last updated: 2026-11-18_

## What is Heretic / ARA and how is it different from plain abliteration?

**Heretic** is a popular open-source tool that automates and refines the abliteration process. Many high-quality community builds (especially from DavidAU, 0bserverx, and others) use Heretic.

**ARA** (often seen as “Heretic + ARA”) is an additional refinement step that further reduces residual refusals while trying to keep the model’s original capabilities intact.

Plain abliteration is the core technique. Heretic and ARA are improved, more automated, or multi-stage versions of it. Builds labeled “Heretic” or “Heretic-ARA” are generally considered more thoroughly processed than early simple abliterations.

_Tags: basics, heretic, methodology · Last updated: 2026-11-15_

## What hardware do I need to run popular abliterated models?

Rough practical guidelines (Q4_K_M or similar quant):

- **8–9B models** → 6–8 GB VRAM (or unified memory)
- **12–14B** → 8–10 GB
- **27B dense** → 16–18 GB
- **35B-A3B MoE** → ~21 GB
- **70B** → 40 GB+

These are estimates. Actual usage depends on context length, quantization, and whether you offload layers to CPU/RAM. Always leave some headroom for the context window.

Mac users with high unified memory (especially 24 GB+) can often run larger models than the VRAM numbers suggest.

_Tags: hardware, vram, setup · Last updated: 2026-11-20_

## What’s the easiest way to run an abliterated model locally?

Most people start with one of these:

1. **Ollama** (simplest)
   - `ollama run huihui_ai/Qwen3.8-abliterated:27b`
   - or `ollama run hf.co/huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF:Q4_K_M`

2. **LM Studio** (best GUI)
   - Download the GGUF from Hugging Face and load it directly.

3. **llama.cpp** (most control)
   - Best for advanced users who want fine-grained settings.

Ollama is usually the fastest path to a working chat. LM Studio is excellent if you prefer a visual interface and easy model management.

_Tags: setup, ollama, lm-studio · Last updated: 2026-11-19_

## Does abliteration destroy model quality?

It depends on the implementation.

Well-done abliteration (especially newer Heretic / ARA builds) usually causes only a small drop in general capability while significantly reducing refusals. Poorly done abliteration can introduce repetition, reduced coherence, or other artifacts.

Many of the current top downloads (huihui-ai Qwen 3.8/3.6, certain DavidAU and 0bserverx builds) are considered to retain strong performance. The best approach is to test the specific model you care about on tasks that matter to you.

_Tags: quality, performance · Last updated: 2026-11-17_

## How can I test if a model is actually low-refusal?

There is no single perfect test, but common approaches include:

- Asking a range of borderline or normally refused questions and measuring how often the model refuses or heavily hedges.
- Using public refusal evaluation sets (some model cards publish their own scores).
- Comparing the same prompts against the original base model.

Be aware that “0 refusals” claims on model cards are often measured on a specific, limited set of prompts. Real-world behavior can still vary.

_Tags: testing, evaluation · Last updated: 2026-11-12_

## Which quantization should I use?

General recommendations:

- **Q4_K_M** or **Q5_K_M** → Best balance of quality and size for most people
- **Q3_K_M / IQ3** → When you are tight on VRAM
- **Q6_K / Q8_0** → When you have spare VRAM and want maximum quality
- **UD-Q4_K_XL** or similar “immatrix” / upgraded quants → Often preferred when available

Start with Q4_K_M. Only go lower if the model doesn’t fit, or higher if you have clear quality issues.

_Tags: quantization, gguf, setup · Last updated: 2026-11-16_

## My abliterated model keeps repeating itself. How do I fix it?

Repetition is a known occasional side effect with some abliterated models.

Common fixes (in order of trying):

1. Increase `repeat_penalty` (try 1.15 → 1.3)
2. Increase `repeat_last_n` (try 512 → 1024 or higher)
3. Lower temperature slightly
4. Use a different quant or a different build of the same base model
5. Ensure you are using the correct chat template

Many users find that a combination of `repeat_penalty 1.25–1.3` + higher `repeat_last_n` solves most looping issues.

_Tags: troubleshooting, repetition · Last updated: 2026-11-18_

## Is it legal to download and run abliterated models?

Downloading and running open-weight models locally for personal use is generally legal in most jurisdictions. The weights themselves are usually distributed under the same license as the base model (Apache-2.0, MIT, Llama license, etc.).

However:

- You are fully responsible for how you use the model.
- Generating or distributing illegal content remains illegal regardless of the model.
- Commercial or high-risk use may trigger additional obligations under regulations such as the EU AI Act.

This is not legal advice. Laws vary by country and change over time.

_Tags: legal, responsibility · Last updated: 2026-11-10_

## Why run an uncensored model locally instead of using an API?

The main advantages of local inference:

- **Privacy** — Your prompts never leave your machine.
- **No usage limits or per-token cost** after the hardware is paid for.
- **Full control** — You decide the system prompt, sampling settings, and whether any safety layer is present.
- **Offline capability**

The trade-off is that you need sufficient hardware and you are responsible for the model’s behavior.

_Tags: privacy, local-vs-api · Last updated: 2026-11-14_

## The model gives weird or broken replies. Is it the chat template?

Yes — incorrect chat templates are one of the most common causes of poor results.

Always check the model card for the recommended template. Popular formats include ChatML, Llama-3, Gemma, and Qwen-specific templates.

In Ollama you can often set it with a Modelfile. In LM Studio and most UIs there is a template selector. Using the wrong template can make even a strong model look broken.

_Tags: troubleshooting, template · Last updated: 2026-11-13_

## Which abliterated model should I try first?

Good starting points in late 2026:

- **Lower VRAM (≤ 8–10 GB)**: DavidAU’s Qwen3.5-9B Heretic builds or Gemma 4 12B abliterated
- **Sweet spot (16–18 GB)**: huihui-ai Qwen3.8-27B-abliterated or strong Heretic 27B variants
- **Higher end**: Qwen3.6-35B-A3B abliterated MoE

Start with a well-known, high-download build from huihui-ai, DavidAU, or 0bserverx. They tend to be the most battle-tested.

_Tags: recommendations, getting-started · Last updated: 2026-11-20_

## How often do new abliterated versions appear?

Very quickly. For major model releases (new Qwen, Gemma, DeepSeek, GLM, etc.), abliterated and Heretic versions often appear within hours to a couple of days.

The ecosystem moves fast. Checking Hugging Face for the base model name + “abliterated” or “heretic” is usually the fastest way to find the latest builds.

_Tags: ecosystem, updates · Last updated: 2026-11-19_

## Should I use Ollama or LM Studio?

**Use Ollama if** you want the simplest terminal/CLI experience, easy scripting, and good integration with other tools.

**Use LM Studio if** you prefer a polished graphical interface, easy model downloading/management, and built-in server mode.

Many people use both. Ollama for daily driving and automation, LM Studio when they want to quickly try different models or adjust settings visually.

_Tags: setup, tools · Last updated: 2026-11-17_