Best Uncensored Local LLMs in 2026: Abliterated and Heretic Models by VRAM

Ask a mainstream chatbot to help with a murder-mystery plot, a gritty war scene, a villain’s monologue or a question about how a piece of malware works so you can defend against it, and there is a fair chance you get a polite refusal. These false refusals are the main reason people look for “uncensored” models. In 2026 the open-model community has made this easy: almost every popular open model now has an abliterated or Heretic version that answers instead of lecturing, and many of them run on a normal gaming GPU. This guide lists the best uncensored local LLMs as of November 2026, sorted by VRAM, with the exact Hugging Face repos, which GGUF file to download, and how to run each one with Ollama and llama.cpp. It also explains the difference between abliteration, Heretic and fine-tunes, how much quality you lose, and how to use these models responsibly.

What “Uncensored” Means for a Local LLM

Instruction-tuned models from Google, Alibaba, DeepSeek and others are trained to refuse a wide range of requests. Some of those refusals are sensible, but many are false positives: a horror story gets flagged as violence, a pharmacology question as drug advice, a penetration-testing lab exercise as hacking. On a cloud service you cannot change that. On your own PC you can.

An “uncensored” model in the local-AI world is simply a version of an open model whose refusal behaviour has been reduced or removed. It does not have new knowledge and it is not smarter than the original. It just stops saying “I can’t help with that” for requests the base model was trained to decline. That makes it more useful for fiction, roleplay, research and red-team work, and it also means you are now the one responsible for what it is used for. There is a short section on that near the end.

Everything below runs fully offline. Nothing you type leaves your machine, which is the other big reason people choose local models.

Abliteration vs Heretic vs Fine-Tunes

You will see three kinds of uncensored model on Hugging Face. They behave differently, so it is worth knowing which one you are downloading.

Abliteration

Researchers found that refusals in most chat models are controlled by a single “refusal direction” inside the network. Abliteration measures that direction by comparing the model’s internal activations on prompts it refuses and prompts it answers, then removes it from the weights. No training data is needed and the rest of the model is barely touched. The best-known publisher is huihui-ai, who releases an abliterated copy of almost every major open model within days of launch. Repos are named Huihui-<model>-abliterated.

Heretic

Heretic by p-e-w automates abliteration. It searches for the ablation settings that cut refusals the most while keeping the model as close as possible to the original, measured as KL divergence on harmless prompts. Its newer Arbitrary-Rank Ablation (ARA) method is now the default. Heretic 2.0 was released on November 1, 2026, with ARA applied as a low-rank update, a plugin system and a BERT-based refusal classifier. It is a breaking update, so 1.x config files no longer work. Heretic models usually publish their refusal count and KL divergence on the model card, which makes them easy to compare. Install it with pip install -U heretic-llm and run heretic <model-id> if you want to make your own.

Fine-tunes

Older “uncensored” models such as Dolphin and many roleplay models were retrained on datasets without refusals. Fine-tunes change the model’s writing style and personality as well, which can be good for creative writing but often costs some reasoning ability. Many 2026 community models combine both: a fine-tune for style plus a Heretic pass for refusals (DavidAU’s “Fable Fusion” models are an example).

Balanced vs Aggressive

Some publishers, notably HauhauCS, release two strengths. Balanced keeps the model’s reasoning habits and is recommended for most work. Aggressive removes more refusals but may be slightly less stable. Start with Balanced and switch only if it still deflects requests you consider legitimate.

Quick Picks by VRAM

Rule of thumb: pick a file 1.5 to 3 GB smaller than your VRAM so the context (KV cache) fits too. Mixture-of-experts models (the ones with “A3B” or “A2.5B” in the name) can go beyond that by keeping some experts in system RAM.

All Models Compared

The Best Uncensored Models in Detail

1. Huihui Qwen3.8 27B abliterated (best overall)

Qwen3.8 27B is the strongest open model that fits a single consumer GPU, and huihui-ai’s abliterated version is the most downloaded uncensored model on Hugging Face right now, with more than 3 million downloads of the GGUF repo. Only the middle layers (22 to 52) are ablated, and the vision encoder and MTP (multi-token prediction) head are left untouched, so it still reads images and still supports fast speculative decoding.

• Repo: huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF
• Ollama: ollama run huihui_ai/Qwen3.8-abliterated (18 GB, 256K context, text and image)
• Files: UD-IQ4_XS 14.4 GB (16 GB cards), Q4_K_M 16.6 GB, UD-Q4_K_XL 17.4 GB, UD-Q5_K_XL 20.7 GB, Q6_K 22.4 GB, Q8_0 29 GB
• Note: in files below Q8_0, the ablated tensors are kept at Q8_0 and the file gets an _L suffix, so a Q2_K_L can be bigger than a normal Q3. That is expected.

Popular alternatives on the same base: 0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF (Heretic plus abliteration) and HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF.

2. Gemma 4 12B QAT Uncensored, HauhauCS Balanced (best for 8 to 12 GB)

Google’s Gemma 4 12B is the sweet spot for mid-range cards, and HauhauCS built this version from Google’s official quantization-aware-trained (QAT) weights, so the single Q4_K_M file loses very little quality. The author reports 0 refusals out of 465 test prompts, while noting that a few edge cases may deflect on the first ask. It keeps vision through a small mmproj file and ships an MTP drafter that the author says speeds up generation by about 60% in llama.cpp.

• Repo: HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced
• Files: ...Balanced-Q4_K_M.gguf (6.9 GB), mmproj-...-BF16.gguf (168 MB, vision), mtp-gemma-4-12B-it.gguf (242 MB)
• Settings from the card: temperature 0.6, top_k 64, top_p 0.9, min_p 0.05, repeat penalty 1.1, context up to 256K

Alternative: culturerevolt/gemma-4-12b-heretic-abliterated-GGUF (about 200K downloads).

3. Gemma 4 31B uncensored Heretic (best for quality on 24 GB)

If you want the bigger Gemma, llmfan46’s Heretic build is one of the most transparent options. The card shows refusals cut from 99/100 to 10/100, a KL divergence of just 0.054, and MMLU only moving from 86.5% to 85.9%. That is a good illustration of how little quality a careful Heretic pass costs. Only the attention output projections in layers 30 to 48 were changed.

• Repo: llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF
• Files: Q4_K_M 18.7 GB, Q5_K_M 21.8 GB, Q6_K 25.2 GB, Q8_0 32.6 GB, plus a BF16 mmproj for images
• Thinking mode: add <|think|> to the system prompt

For 16 GB, the mixture-of-experts sibling mradermacher/gemma-4-26B-A4B-it-ultra-uncensored-heretic-i1-GGUF (220K downloads) is faster, with only about 4B active parameters per token.

4. Huihui Qwen3.6 35B-A3B abliterated MTP (fast MoE for 16 to 24 GB)

Qwen3.6 35B-A3B activates only 3B parameters per token, so it generates text far faster than a dense 27B model. On 16 GB you keep the attention layers on the GPU and push some expert weights to system RAM. huihui-ai’s MTP build keeps the multi-token-prediction head for even more speed and has more than 2.2 million downloads.

• Repo: huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF
• Files: Q2_K 13.2 GB, Q4_K, Q6_K 29.2 GB, Q8_0 37.8 GB
• llama.cpp flag for MTP: --spec-type draft-mtp

Popular alternative: HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive (2.4 million downloads).

5. Gemma 4 E4B Uncensored, HauhauCS Aggressive (laptops and 6 GB)

Gemma 4 E4B is a 4B-effective model built for phones and laptops, and it understands images and audio as well as text. The HauhauCS version is one of the most downloaded uncensored models of the year (2.3 million). Q4_K_M is 5.0 GB and IQ4_XS is 4.8 GB; add the 945 MB mmproj file for image and audio input. Use temperature 1.0, top_p 0.95, top_k 64 and the --jinja flag in llama.cpp.

• Repo: HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive
• Alternative: llmfan46/gemma-4-E4B-it-uncensored-heretic

6. Flint 40B-A2.5B abliterated (new: coding and security research)

Recursive released Flint on October 21, 2026: a 40B mixture-of-experts coding model with only 2.5B active parameters, 256 experts, a 262K context and an Apache 2.0 licence. Recursive reports 74.2 on SWE-bench Verified. Within two days dima6312 published an abliterated version aimed at defensive security work, where base models often refuse to explain attack techniques that analysts need to understand. The card is unusually thorough: refusals on 100 held-out MITRE ATT&CK prompts fell from 15% to 0%, false refusals stayed at 0%, KL divergence is 0.0146, and no significant drop was found on HumanEval+, MBPP+, GSM8K or MMLU-Pro.

• Repo: dima6312/Flint-40B-A2.5B-abliterated-GGUF (imatrix quants: mradermacher/Flint-40B-A2.5B-abliterated-i1-GGUF)
• Files: IQ4_XS 21.3 GB, MXFP4_MOE 22.2 GB, Q4_K_M 24.2 GB, Q5_K_M 28.3 GB, Q8_0 42.4 GB
• Needs llama.cpp b11827 or newer. Recommended sampling: temperature 1.0, top_p 0.95, top_k 20.
• On 24 GB it fits only with a short context; it is happiest on 32 GB cards or 64 GB Macs.

7. Creative-writing fine-tunes: Fable Fusion and Dolphin Venice

If your main use is fiction and roleplay, a fine-tune often writes better prose than a plain abliterated model. Two popular choices:

• DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF: a storytelling-focused Qwen3.6 27B with a Heretic pass, and the single most downloaded Heretic GGUF (over 2.5 million). Same VRAM needs as any 27B model. For 8 GB, DavidAU’s Qwen3.5 9B Defiant Fable is the smaller sibling.
• Dolphin Mistral 24B Venice Edition: an older (2025) fine-tune that is still widely used because it follows the system prompt closely and has a neutral tone. GGUFs: bartowski.

8. Big MoE models for high-RAM PCs and workstations

These do not fit in VRAM, but with 96 GB or more of system RAM, llama.cpp can keep the experts in RAM and the rest on the GPU.

• Huihui Qwen3.8-Flash-Next abliterated (125B total, 6B active, image input): IQ2_XS 68.2 GB, IQ3_XXS 76 GB, UD-IQ4_XS 93.7 GB, UD-Q4_K_XL 111 GB. Repos: v1 and v2 (Oct 11). Its 6B active size makes it surprisingly quick for its quality.
• Huihui GLM-5.3-Flash abliterated (320B, 18B active, MIT): UD-IQ2_XXS 102 GB, UD-IQ4_XS 157 GB. Repo: huihui-ai/Huihui-GLM-5.3-Flash-abliterated-GGUF.
• Huihui DeepSeek V4.1 Flash 1017 abliterated (DeepSeek’s October 17 update, MIT): needs llama.cpp b11870 or newer. Repo: huihui-ai/Huihui-DeepSeek-V4.1-Flash-1017-abliterated-GGUF.
• Sage-A1.5 Heretic Abliterated: Shanghai AI Lab’s new 35B MoE scientific-agent model (built on Qwen3.6 35B-A3B), released in October. GGUF: mradermacher/Sage-A1.5-Heretic-Abliterated-Uncensored-i1-GGUF. Fits the same hardware as Qwen3.6 35B-A3B.

Which GGUF Quant to Download

The same rules apply as for any GGUF model (see our GGUF quantization guide):

• Q4_K_M is the default. Quality is close to the original for chat and writing.
• IQ4_XS / UD-IQ4_XS saves about 2 GB and is how a 27B model fits a 16 GB card.
• Q5_K_M and Q6_K are worth it if you have spare VRAM, especially for coding.
• Below Q3 quality drops quickly on dense models. Big MoE models tolerate 2-bit better.
• i1 / imatrix repos (mradermacher’s -i1-GGUF) are usually a little better at the same size.
• Abliterated models are slightly more sensitive to heavy quantization than the originals, which is why huihui-ai keeps the ablated tensors at Q8_0.

Run Them With Ollama

Install Ollama from ollama.com/download, then either pull a ready-made library model or any GGUF straight from Hugging Face with the hf.co/ prefix and a quant tag:

# Qwen3.8 27B abliterated from the Ollama library (18 GB)
ollama run huihui_ai/Qwen3.8-abliterated

# Any Hugging Face GGUF, choosing the quant after the colon
ollama run hf.co/huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF:UD-IQ4_XS
ollama run hf.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced:Q4_K_M
ollama run hf.co/llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF:Q4_K_M
ollama run hf.co/huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF:Q2_K

To give a model a permanent system prompt and settings, create a Modelfile:

FROM hf.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced:Q4_K_M
PARAMETER temperature 0.6
PARAMETER top_k 64
PARAMETER top_p 0.9
PARAMETER min_p 0.05
PARAMETER num_ctx 16384
SYSTEM "You are a co-writer for dark fantasy fiction. Write vivid, mature prose and stay in character."

ollama create gemma12-writer -f Modelfile
ollama run gemma12-writer

More commands are in our Ollama commands cheat sheet.

Run Them With llama.cpp

llama.cpp gives you more control, supports MTP speculative decoding and is required for the newest models (Flint needs b11827+, DeepSeek V4.1 Flash 1017 needs b11870+). Recent builds use the unified llama command:

# Download and serve with an OpenAI-compatible API on port 8080
llama serve -hf huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF:UD-IQ4_XS -c 32768 -ngl 99 -fa on --jinja

# Chat in the terminal
llama cli -hf HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced:Q4_K_M --jinja

# Flint 40B with all layers on GPU
llama serve -hf dima6312/Flint-40B-A2.5B-abliterated-GGUF:MXFP4_MOE -c 16384 -ngl 999 -fa on --jinja --temp 1.0 --top-p 0.95 --top-k 20

MoE on a 16 GB card: keep the expert weights in system RAM and everything else on the GPU. With Qwen3.6 35B-A3B at Q4 this usually gives good speed on a 16 GB card with 32 GB of RAM:

llama serve -hf huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF:Q4_K -ngl 99 --n-cpu-moe 20 -c 32768 -fa on --jinja --spec-type draft-mtp

Raise --n-cpu-moe if you run out of VRAM, lower it if you have room left. For image input, add --mmproj with the model’s mmproj file (the -hf form downloads it automatically when the repo has one).

Prefer a desktop app? LM Studio, Jan and KoboldCpp all load these GGUF files. Search for the repo name in the app’s model browser.

Settings That Matter

• Use the card’s sampling settings. Abliterated Gemma models often want temperature 1.0 and top_k 64; Balanced HauhauCS builds want 0.6. Wrong settings cause repetition or rambling more often than the uncensoring does.
• Write a clear system prompt. An uncensored model is not automatically a better writer. Tell it the genre, tone, point of view and what it should not do (for example “do not add disclaimers”).
• Turn thinking off for chat and fiction. Reasoning modes add long hidden text and slow replies. Keep them on for coding and research.
• Keep context realistic. 16K to 32K tokens is plenty for most chats and keeps VRAM free. Set num_ctx in Ollama or -c in llama.cpp.
• Use MTP files when offered. Speculative decoding gives a noticeable speed-up with the same output.

Good Uses for Uncensored Models

Responsible Use

Removing refusals removes a safety layer, and the model’s publishers say so on every card: huihui-ai recommends research or controlled use and warns against public-facing or production deployments. A few practical rules:

• The law still applies. Using a model to plan harm, harass people, create non-consensual content about real people or break computer-misuse laws is illegal whatever the model is willing to write.
• Do not put an uncensored model behind a public chatbot without your own moderation layer. Keep Ollama and llama.cpp servers bound to localhost, or protect them with authentication.
• Keep them away from children. Use a normal aligned model on shared family PCs.
• Check facts. An uncensored model is as likely to be wrong as the original, and it will not warn you. Verify medical, legal and security information.
• Respect licences. Abliterated models inherit the base licence (Apache 2.0 for Qwen and Gemma 4 31B, Gemma terms for HauhauCS’s Gemma builds, MIT for GLM and DeepSeek, Qwen Community 1.0 for Qwen3.8-Flash-Next).

Troubleshooting

• “unknown model architecture” or the file will not load: update llama.cpp, Ollama or LM Studio. New bases like Flint and DeepSeek V4.1 Flash need very recent builds.
• The model still refuses sometimes: rephrase or regenerate, set a system prompt that states the context (for example a fiction project), or try the Aggressive variant. No uncensored model reaches zero on every prompt.
• Gibberish or endless repetition: usually a missing chat template. Add --jinja in llama.cpp and use the card’s sampling settings.
• Out of memory: drop one quant level, lower the context, or for MoE models raise --n-cpu-moe.
• Ternary or “Bonsai” files fail on stock llama.cpp: some repos need a special fork. Stick to normal GGUF quants unless you know you need them.

FAQ

What is the best uncensored LLM to run locally in 2026?

For a 24 GB GPU, Huihui Qwen3.8 27B abliterated. For 16 GB, the same model at UD-IQ4_XS or Qwen3.6 35B-A3B abliterated with expert offload. For 8 to 12 GB, Gemma 4 12B QAT Uncensored. For laptops, Gemma 4 E4B Uncensored.

What does abliterated mean?

The model’s internal “refusal direction” has been found and removed from its weights, so it stops refusing requests without being retrained.

Is Heretic better than abliteration?

Heretic is an automated, optimised form of abliteration. It tunes the settings to remove refusals while keeping the model closer to the original, and it reports refusal and KL-divergence numbers, so Heretic models are easier to compare.

Do uncensored models lose quality?

Good ones lose very little. The Gemma 4 31B Heretic build drops from 86.5% to 85.9% on MMLU, and the Flint abliteration shows no significant drop on coding and maths benchmarks. Crude abliterations and heavy fine-tunes can lose more.

Can I run an uncensored model in Ollama?

Yes. Use ollama run huihui_ai/Qwen3.8-abliterated, or load any Hugging Face GGUF with ollama run hf.co/<user>/<repo>:<quant>.

Are uncensored LLMs legal?

Downloading and running open-weight models is legal in most countries, and the licences allow modification. What you do with the output is still subject to the law.

Which uncensored model is best for creative writing?

DavidAU’s Qwen3.6 27B Fable Fusion Heretic for 24 GB, his Qwen3.5 9B Defiant Fable for 8 GB, and Dolphin Mistral 24B Venice Edition if you prefer a neutral, instruction-following style.

Final Thoughts

Uncensored local models have gone from a niche hack to a routine release: every major open model now gets abliterated or Heretic versions within days, and the quality cost of a careful pass is close to zero. Pick by VRAM: Qwen3.8 27B abliterated on 16 to 24 GB, Gemma 4 12B Uncensored on 8 to 12 GB, Gemma 4 E4B on laptops, and Flint 40B if you do security work. Use the card’s settings, write a clear system prompt, keep the server private, and you get a model that helps with fiction, research and hard questions instead of refusing them.

Related guides:

• GGUF Quantization Explained: Which File to Download
• Ollama Commands Cheat Sheet
• Best Local LLMs With Vision and Video Understanding
• Liquid AI LFM3: Run It Locally
• Kimi K3.1 Pro: Can You Run It Locally?
• How to Run Recursive Flint Locally on a 16 GB GPU
• Uncensored Local AI Image Generation in ComfyUI
• Uncensored Local AI Video Generation in ComfyUI

Sources

• huihui-ai Qwen3.8 27B abliterated card and Ollama page
• HauhauCS Gemma 4 12B and Gemma 4 E4B cards
• llmfan46 Gemma 4 31B Heretic card
• Flint 40B-A2.5B abliterated card and RunAIHome: Recursive Flint 40B hardware guide
• TechNode: DeepSeek V4.1 Flash 1017
• GitHub: p-e-w/heretic and Heretic 2.0 release report
• Qwen3.8-Flash-Next abliterated, GLM-5.3-Flash abliterated and Qwen3.6 35B-A3B abliterated cards
• InternScience Sage-A1.5 card
• MorphLLM: Best open source LLMs (October 2026) for base-model release dates
• Hugging Face model search, sorted by downloads and by date, for “abliterated”, “heretic” and “uncensored” (checked November 2, 2026)