Best Uncensored Local LLMs by VRAM Tier (2026)

📚 Related: VRAM Requirements for Local LLMs · Qwen Models Guide · Mistral & Mixtral Guide · Best LLMs for Writing

Ask a standard instruct model to write a scene where a character dies on-page and you’ll get a polite refusal. Ask it to walk through the chemistry of explosives for a novel and it’ll lecture you about safety. Ask it to roleplay a villain and it’ll break character mid-monologue to remind you that violence is bad.

Those refusals exist because instruct-tuned models are trained to refuse broad categories of content. That training serves customer-facing chatbots well. For fiction writers, security researchers, philosophers, roleplay enthusiasts, and anyone whose work involves the full range of human experience, it’s just in the way.

Uncensored models fix this. What’s changed in the four months since this guide first ran: the landscape moved hard. Qwen 3.6 abliterated and the new Gemma 4 Heretic builds replaced most of what used to be the top of this list. The Llama 3.1 and Qwen 2.5 era picks still work, but they’re a generation behind. Here’s the current curated lineup, organized by what fits in your VRAM.

What “Uncensored” Actually Means

Standard instruct models go through alignment training: reinforcement learning from human feedback (RLHF) or similar techniques that teach the model to refuse certain categories of requests. That’s where “I can’t help with that” comes from.

Uncensored models have that alignment layer removed. Three techniques dominate in mid-2026:

Dataset filtering (Dolphin method). Eric Hartford and the Cognitive Computations team filter the training dataset to remove refusal, avoidance, and hedging examples before fine-tuning. The model never learns to refuse because refusals aren’t in the training data. This produces a model that engages with any topic without the hedging.

Abliteration. A technique that removes alignment from an already-trained instruct model. Refusal behavior turns out to live in a single direction in the model’s activation space — a specific mathematical vector. Abliteration identifies that direction and subtracts it from the weights, producing an uncensored model without retraining. This is faster and cheaper than dataset-filtered fine-tuning, and it preserves the base model’s behavior on everything that isn’t refusal-adjacent. Maxime Labonne’s abliteration tutorial is the canonical write-up.

Heretic (Gemma 4 era). A refined, automated abliteration approach that emerged for Gemma 4 specifically, because Gemma 3 turned out to be unusually resistant to standard abliteration. Instead of a single hand-tuned ablation, Heretic searches for the ablation that strips out the most refusals while diverging least from the original — it optimizes those two quantities directly (refusal count and distance from the base model). The result for Gemma 4 Heretic builds: most refusals removed while normal-prompt behavior stays very close to the base.

What uncensored does NOT mean:

• Not “jailbroken.” Jailbreaking tricks a model into ignoring alignment through prompt engineering. That’s fragile and inconsistent. Uncensored models have the alignment genuinely removed at the weights level — they’re not fighting their training, the training is changed.
• Not “will help you do crimes.” Removing alignment doesn’t add knowledge. The model still has the same training data and the same cutoff. It won’t teach you things it doesn’t know.
• Not “no limitations.” Hallucination, knowledge cutoff, context drift, reasoning failures — all still present. The model just won’t refuse to try.

Why You’d Want One

If standard models never refuse your requests, uncensored models wouldn’t exist. But alignment is a blunt instrument that blocks legitimate work alongside actually harmful requests.

Fiction writing. A thriller needs a violent scene. Literary fiction has sexual content. A war novel needs the war. Standard models refuse or sanitize these into something unusable. The uncensored variants write what you ask for.

Academic research. Studying radicalization, analyzing propaganda techniques, researching drug pharmacology, dissecting an extremist manifesto — standard models treat the request as the act. Uncensored variants engage with the material.

Red teaming and security testing. Testing AI defenses requires adversarial content. You can’t pressure-test a refusal classifier with a model that itself refuses.

Roleplay and interactive fiction. Tabletop RPG scenes, character-driven roleplay, interactive stories — standard models break character to insert safety disclaimers. Uncensored variants stay in character.

Philosophy and ethics. Steelmanning a morally uncomfortable position without hedging is what philosophy actually requires. Standard models hedge. Uncensored ones engage.

Just running your own hardware on your own terms. Some people don’t want a corporation’s content policy deciding what their local AI will discuss. That’s a defensible position.

By VRAM Tier

Pick the tier that matches your hardware. Every model below has been confirmed live on HuggingFace at the time of writing, with GGUF available either in the linked repo or via a sibling quant repo from bartowski or mradermacher.

8GB VRAM — Budget tier

The 8B class is the floor for usable uncensored output. Picks:

If you’re choosing between these three on 8GB and don’t have a specific reason, Qwen3-8B abliterated v2 is the most current pick. Dolphin 3.0 Llama 3.1 8B is the proven creative-writing pick. The Llama 3.1 abliterated GGUF is the “just works” option.

12GB VRAM — Headroom tier

12GB lets you run 8B class at higher quants (Q6 or Q8 for noticeably better output quality) or step up to the 12B tier:

The Q2_K trick is fragile. If you have 12GB and want the 27B base, you’re probably better off saving for a 16GB card than running 27B at Q2.

16GB VRAM — Sweet spot tier

This is where the genuinely current 24B-class models become viable:

For most 16GB rigs the choice is between Gemma 4 26B-A4B Heretic (MoE, faster, vision-capable) and Dolphin 3.0 Mistral 24B (dense, the established uncensored all-rounder). The two represent different philosophies (fresh-architecture vs proven-finetune), and either is defensible.

24GB VRAM — The flagship tier (RTX 3090 / 4090)

This is where the most current picks live. All four below are confirmed live as of July 2026:

If forced to pick one: Qwen 3.6-27B abliterated for general work, Gemma 4 31B Heretic if you need vision or tool calling, Dolphin 3.0 Mistral 24B if you prefer a dataset-filtered base over an abliterated one.

There’s also a coding-specific variant worth knowing about: huihui-ai/Huihui-Qwen3-Coder-Next-abliterated — an abliterated coder variant for people who specifically want uncensored code assistance.

48GB+ VRAM — Dual GPU / pro tier

This tier means dual 3090s, a 4090 + 3090, or workstation cards. The picks change shape — you can run actual 70B-class models or quantized MoE frontier weights:

For most readers this tier means R1-Distill 70B abliterated. V4 Flash abliterated is for people running rented or owned datacenter hardware.

What About Llama 4?

Llama 4 (Scout and Maverick) shipped in April 2026 — see the Llama 4 guide for the architecture. This is the part that’s changed since the guide first ran: a fully-abliterated Llama 4 Scout (17B active, 16-expert MoE) has now surfaced, with GGUF quants mirrored by mradermacher. The abliteration itself is a community upload rather than one of the huihui/mlabonne mainline accounts, so treat it as less battle-tested than the Qwen and Gemma picks. Maverick still has no abliterated build. So Scout is now an option if you specifically want uncensored on Llama 4 — but the center of gravity is still firmly in Qwen 3.6 and Gemma 4 territory, which is where the abliteration community is putting its effort.

Where to Find Them

HuggingFace

The HuggingFace repos above link directly. For finding additional variants:

• huihui-ai — the dominant abliterated-model uploader in mid-2026. Maintains current abliterations for Qwen 3.6 / 3.5, DeepSeek V4 Flash, DeepSeek 671B, and most major open releases.
• mradermacher — reliable quantizer with broad coverage; current source for most Gemma 4 Heretic GGUF mirrors. Offers both static and imatrix quants.
• bartowski — current go-to for high-quality GGUF quants across the catalog. Uses imatrix quantization for better quality at low bit rates.
• mlabonne — author of the canonical abliteration tutorial; maintains a collection of abliterated models including Gemma 3 27B abliterated, which is still in active use.
• dphn — Cognitive Computations / Eric Hartford’s official Dolphin uploads. The line has since moved past 3.0 to the Dolphin X1 series (X1-8B, the tiny X1 Trinity Nano MoE, and the Mistral-24B Venice Edition). Dolphin 3.0 is still the proven pick in this guide, but X1 is where new work is landing.

Ollama

For models with Ollama community uploads, the pattern is ollama pull huihui_ai/<model-name> for huihui-ai’s catalog:

ollama pull huihui_ai/Qwen3.6-abliterated
ollama pull huihui_ai/qwen3.5-abliterated

For models without an Ollama upload, the standard pattern is: download the GGUF from HuggingFace, then import via Modelfile.

# After downloading the GGUF
echo 'FROM ./model-name.Q4_K_M.gguf' > Modelfile
ollama create my-uncensored-model -f Modelfile
ollama run my-uncensored-model

The Quality Tradeoff

Uncensored models aren’t a free lunch. Removing alignment has a cost.

Abliteration preserves most capability because it targets a narrow mathematical direction. On normal, non-refused prompts, the model behaves identically to the base. On edge cases (prompts near the refusal boundary), you may see slightly degraded coherence. Heretic builds (for Gemma 4) refine this further by automatically tuning the ablation to strip refusals while keeping the model statistically very close to the base on standard outputs.

Dataset-filtered finetunes (Dolphin) are a different model. The training data is different, so style, instruction-following, and knowledge retention can drift. Dolphin 3.0 is well-tuned, but it’s not the same as “base instruct minus refusals.” It’s its own thing.

Practical impact:

• For creative writing, the tradeoff is almost always worth it. The quality cost is small; the freedom to write without refusals is large.
• For coding, the base instruct model usually wins unless it’s actively refusing something. If you need uncensored and code, the new Qwen3-Coder-Next abliterated variant is more directly purpose-built than asking Dolphin to code.
• For reasoning, abliterated versions of reasoning models (DeepSeek R1 distill, Qwen 3.5 reasoning) preserve reasoning quality well. Dolphin 3.0 R1 Mistral 24B is purpose-built for “reasoning + uncensored” combined.

Test before committing. Run the same prompts through the base instruct and the uncensored variant. If the uncensored version handles your actual workflow, use it. If you see quality drops on the tasks you care about, try a different uncensored variant or a different technique.

The Ethics of Uncensored Models

This is straightforward: you own your hardware, and you control what runs on it.

When Meta, Mistral, Alibaba, Google, or DeepSeek release open-weight models, they’re giving you the weights under a license that permits modification. Fine-tuning to remove alignment is no different from fine-tuning for any other purpose — it’s using the model exactly as the license allows.

The alignment in standard models reflects one company’s content policy, designed for that company’s liability when serving millions of users through a web API. Those concerns don’t apply to a model running on your own hardware for your own work. A fiction writer doesn’t need OpenAI’s content policy. A security researcher doesn’t need Meta’s refusal training. A philosopher doesn’t need Google’s hedging.

With that freedom comes responsibility. An uncensored model will do what you ask. It won’t remind you about ethics, suggest you reconsider, or refuse to engage. The judgment is entirely yours. For most people running local AI for creative and research work, that’s exactly how it should be.

What Uncensored Won’t Do

A few misconceptions worth clearing up:

It won’t teach you things the model doesn’t know. Removing alignment doesn’t add knowledge. If the training data didn’t include working exploit code, classified intelligence, or specific synthesis procedures, the uncensored model doesn’t have them either. It just won’t refuse to discuss the topic in general terms.

It won’t access the real world. An uncensored model still generates text in its context window. It can’t hack systems, access databases, or interact with anything outside the prompt and the response.

It won’t be uniformly “better.” On most everyday tasks (summarization, Q&A, coding, analysis), the standard instruct model performs the same or slightly better because its alignment training also polishes general response quality. Uncensored is a win specifically when alignment is in your way.

It won’t bypass the model’s actual limits. Hallucination, knowledge cutoff, context drift, reasoning failures — all still present. Uncensored models tend to be more honest about not knowing things, which can actually make it easier to spot when they’re making things up.

Recommended Setups by Use Case

Fiction writing on 8GB: Qwen3-8B abliterated v2. Light hardware, current architecture, no refusals on creative work.

Fiction writing on 16-24GB: Dolphin 3.0 Mistral 24B or Qwen 3.6-27B abliterated. Both excellent for long-form creative work; pick Dolphin for the proven dataset-filter approach, Qwen 3.6 for the current architecture.

Vision + uncensored: Gemma 4 31B Heretic (24GB) or Gemma 4 26B-A4B Heretic (16GB). Native vision and tool calling, abliterated refusals.

Reasoning + uncensored: Dolphin 3.0 R1 Mistral 24B (16-24GB) or DeepSeek-R1-Distill-Llama 70B abliterated (48GB+). The R1-trained reasoning traces survive the dataset filtering / abliteration.

Coding + uncensored: Qwen3-Coder-Next abliterated (24GB) is the most directly purpose-built option. For general coding without restriction issues, the base Qwen 3.6-27B instruct (not abliterated) is usually a stronger coder.

Maximum quality at any cost: DeepSeek V4 Flash abliterated, if you have the hardware to run a 284B MoE. Otherwise R1-Distill 70B abliterated at Q4 on dual 3090s.

The Bottom Line

Uncensored models exist because alignment training blocks legitimate creative and research work alongside genuinely harmful content. If standard models never refuse your requests, you don’t need these. If they do, the current picks are:

• 24GB rig: Qwen 3.6-27B abliterated, or Gemma 4 31B Heretic if you want vision and tools.
• 16GB rig: Dolphin 3.0 Mistral 24B, or Gemma 4 26B-A4B Heretic for MoE speed.
• 8GB rig: Qwen3-8B abliterated v2, or Dolphin 3.0 Llama 3.1 8B for the proven creative pick.
• 48GB+ rig: DeepSeek-R1-Distill-Llama 70B abliterated for reasoning at scale.

The privacy advantages of local AI get even stronger with uncensored models — there’s no content policy log somewhere tracking your generations. Your creative work stays on your machine, unrestricted and unmonitored.

The early-2026 lineup (Llama 3.1 abliterated, Qwen 2.5-based Dolphin) is mostly superseded but still functional if you already have it loaded. Anything new gets started on the current architectures — Qwen 3.6 and Gemma 4 are where the abliteration community is focused now.

Related Guides

• VRAM Requirements for Local LLMs
• Qwen Models Guide
• Qwen 3.6 Local Guide
• Is Qwen Going Closed? Open Weights vs Frontier (2026)
• Best LLMs for Writing & Creative Work
• Mistral & Mixtral Guide
• Llama 4 Guide: Scout and Maverick
• DeepSeek V4 Flash vs Pro
• Running AI Offline
• Local AI Privacy Guide