The 7 Best Uncensored LLMs You Can Run Locally in 2026

The local uncensored model scene changed on August 14, 2026. That day, Alibaba pushed the open weights for Qwen 3.8-27B to Hugging Face, and within hours the community had abliterated builds quantized down to sizes a single RTX 5090 or a 24GB MacBook can hold. For the first time, the strongest model you can run without refusals is also close to the strongest open model, period.

That makes most “best uncensored LLM” lists obsolete. Rankings still circulating in 2026 recommend Llama 2 finetunes and Vicuna, models three generations behind what you can pull from Hugging Face today. This list starts from scratch: seven models, all verified downloadable right now, bucketed by the VRAM you own rather than by benchmark trivia.

One definition before the ranking, because the terms get blurred everywhere. An uncensored LLM is a model whose refusal behavior has been removed or never trained in. It will not decline a prompt on policy grounds. That also means it carries no guardrails at all: you are the safety layer, and outputs are your responsibility. Every model below is a research and self-hosting tool, not a hosted assistant.

How “uncensored” happens: three different methods

Knowing which method produced a model tells you how it will behave.

Abliteration. Researchers identify the internal activation direction responsible for refusals and surgically remove it from the weights. No retraining involved. The model keeps its base intelligence almost intact but stops declining requests. Community builders like huihui-ai and orcarouter publish abliterated versions of major open models within days of release.

Refusal-free finetuning. The Dolphin approach: retrain the base model on a curated dataset with refusals stripped out. This changes the model’s personality more than abliteration does, and quality depends on the finetune dataset.

Neutral alignment. Nous Research trains Hermes models to follow your system prompt instead of a vendor ethics code. The model is not lobotomized; it is steerable. You define the rules per deployment.

All three produce a model that answers where commercial assistants refuse. They differ in how much base capability survives and how much control you keep.

How this list is ranked

Three criteria, in order:

1. Base model strength. An uncensored build inherits the ceiling of its base. A 2026 base beats a 2024 base at the same size.
2. Runs on hardware people own. Every entry lists the quant size and the minimum VRAM or unified memory that holds it. Buckets: 12 to 16GB, 16 to 24GB, and workstation class.
3. Verified availability. Every model links to a live Hugging Face repo or official page. No dead links, no “coming soon.”

Quick comparison before the details:

1. Qwen 3.8-27B Uncensored: the new best overall

The base model is the story here. Qwen 3.8-27B is the open-weight dense sibling of Qwen 3.8-Max, released August 14, 2026 on Hugging Face and ModelScope. It understands images and video natively, targets coding and agent workloads, and posts public benchmark numbers that crowd the closed frontier models. Nothing else on this list has a base this current.

The community moved fast. orcarouter/Qwen3.8-27B-Uncensored-GGUF is the abliterated GGUF build, with quants mapped to real hardware:

• Q4_K_M, 16.8GB: the pick for a 24GB GPU (RTX 3090/4090/5090) or a 32GB Mac
• IQ4_XS, 15.3GB: squeezes onto a 16GB GPU
• Q3_K_M, 13.5GB: when you want context headroom over precision

An FP8 build exists for vLLM if you serve on a workstation card, and a more aggressive variant (0xKitkat/Qwen3.8-27B-Uncensored-Aggressive) trades more refusal-adjacent behavior away at some capability cost. On a desktop RTX 5090, expect around 45 tokens per second at Q4; owners had it running in daily setups within an hour of the weights landing.

One setup gotcha: the qwen35 architecture and MTP support landed in llama.cpp in May 2026. Update to a build from 2026-05 or newer, or the GGUF will not load. Same rule applies through Ollama and LM Studio, which bundle llama.cpp underneath.

Best for: anyone with 16GB+ of VRAM who wants the strongest local model, censored or not. The fact it happens to be uncensored is a bonus on top of a frontier-adjacent base.

2. Dolphin 3.0 Mistral 24B: the finetune veteran, still sharp

Eric Hartford’s Dolphin series is the longest-running uncensored project, and Dolphin3.0-Mistral-24B is its current flagship. Unlike the abliterated entries, Dolphin is a full refusal-free finetune: retrained on a curated dataset, tuned for instruction following, coding, math, and function calling.

Two things keep it on the list in 2026. First, the R1 variant adds reasoning, trained for three epochs on 800k reasoning traces from the Dolphin-R1 dataset, so you get chain-of-thought behavior without a vendor filter deciding which thoughts are allowed. Second, the ecosystem: Dolphin has first-class Ollama support, mature GGUF quants at every size, and years of community prompt patterns.

At 24B dense, Q4 quants land around 14 to 15GB, comfortable on a 24GB card and workable on 16GB with a smaller quant.

Best for: general-purpose local chat and agent work when you prefer a deliberately retrained model over a surgically edited one, and for the R1 build if you want uncensored reasoning.

3. Hermes 4: uncensored by philosophy, not surgery

Hermes 4 from Nous Research takes the position that alignment belongs to the operator. The models are trained to follow your system prompt rather than a company ethics code. Nous calls it neutral alignment, and Hermes 4 tops RefusalBench among open and closed models for complying with user intent without blanket refusals.

The family spans 14B, 36B (including the newer 4.3-36B build), 70B, and a 405B for people with server racks. All are hybrid reasoners: include the reasoning tag and the model thinks longer on hard problems, omit it and you get fast direct answers.

The practical difference from abliterated models matters. An abliterated model cannot refuse anything, ever. Hermes will follow whatever policy you write into the system prompt, including one with limits you choose. For a lot of self-hosters, that is the more useful property: uncensored by default, governable on demand.

Best for: operators who want to define model behavior themselves. The 14B fits a 12GB card at Q4; the 36B wants 24GB.

4. Huihui Qwen 3.6-27B abliterated: the proven workhorse

Before Qwen 3.8 landed, huihui-ai’s Qwen 3.6 abliterations were the default recommendation for local uncensored use, and they remain the safe pick. Qwen 3.6 (April 2026) proved an unusually clean base for abliteration: the refusal direction removes with minimal capability loss, which is why the 27B build shows up at the top of community rankings all year.

The ecosystem is the draw. Huihui publishes the full quant spread on Hugging Face and mirrors to Ollama, so ollama run gets you started in one command. There is a 14B vision-capable variant for smaller cards, and a Samantha personality finetune stacked on the abliteration if you want a warmer conversational default.

Pick this over entry #1 if you value a battle-proven build with months of community validation over the newest base, or if the Qwen 3.8 llama.cpp requirement blocks your current toolchain.

Best for: a stable, widely validated daily driver on 16 to 24GB of VRAM.

5. Gemma 4 26B-A4B uncensored: speed on modest hardware

Most models on this list are dense, so every token pays for every parameter. Gemma 4 26B-A4B is a mixture-of-experts build: 26B total parameters, around 4B active per token. The abliterated version gives you uncensored output at throughput a dense 27B cannot touch on the same card.

That makes it the bucket winner for 12 to 16GB setups. Where a dense 27B at Q3 feels compromised on a 16GB GPU, the A4B architecture keeps quality up while generating multiples faster. On Apple Silicon with 16GB unified memory, it is the difference between usable and painful.

The trade-off is depth on hard reasoning tasks, where dense entries #1 and #2 pull ahead. For summarization, drafting, extraction, and chat, you will rarely notice.

Best for: laptop-class hardware, 16GB Macs, and anyone who values tokens per second over maximum reasoning depth.

6. Mistral Small 3.2 abliterated: the writer’s pick

Ask the roleplay and fiction communities which uncensored model they run and the answer in 2026 is consistently the Mistral Small 3.2 abliteration. Mistral’s base models have a distinct prose quality: less list-happy, better at holding a voice across long context, less prone to the assistant-brained tone that leaks through Qwen and Llama finetunes.

Abliteration matters more for creative writing than for code. Commercial models refuse or sanitize a large share of legitimate fiction: villains, violence, morally gray narrators. An abliterated Mistral Small writes the scene you asked for and keeps the register you set.

At roughly 24B dense, quants land like Dolphin’s: Q4 around 14GB, happy on 16GB cards and up.

Best for: fiction, roleplay, and any writing work where refusal-adjacent sanitization ruins output quality.

7. Huihui GLM-5.1 abliterated: the ceiling

The biggest open abliterated model available: Huihui-GLM-5.1-abliterated-GGUF, a 754B-parameter MoE that ships as a 236GB file even at IQ2_M quantization. This is not a consumer model. It wants a 256GB+ Mac Studio, a multi-GPU rig, or a server with serious RAM for CPU offload.

It earns its slot because it answers a question the other six cannot: how far does uncensored local AI scale? The answer is now “to a model that competes with hosted frontier systems,” which was not true a year ago. If you have the hardware, nothing self-hosted and unrestricted comes close.

Best for: Mac Studio owners, homelab enthusiasts with workstation budgets, and research groups that need frontier-class capability with zero external policy.

Running any of these: serve it, then treat it like an API

Every model above deploys the same way: llama.cpp, Ollama, or LM Studio for GGUF builds, vLLM for FP8. All of them expose an OpenAI-compatible endpoint on localhost, which means your local model is an API the moment it loads:

ollama run huihui_ai/qwen3.6-abliterated:27b
# OpenAI-compatible endpoint now live at http://localhost:11434/v1

That endpoint deserves the same treatment as any API you build against. Point Apidog at http://localhost:11434/v1/chat/completions and you can save prompt payloads as reusable requests, diff responses across quants of the same model, assert on response structure before wiring the endpoint into an app, and mock the model’s responses so your integration tests stop burning GPU time. The workflow is identical to the one in our Kimi K3 local guide: pull weights, serve, then debug the endpoint like a production service. Download Apidog and the whole loop runs offline, which fits the reason you went local in the first place.

Uncensored models make endpoint-level testing more important, not less. With no refusal layer in the model, your application needs its own input and output checks, and those are exactly the request and response assertions an API client automates.

FAQ

Is it legal to download and run these models? Downloading open-weight models and abliterated derivatives from Hugging Face is legal in most jurisdictions. Each repo carries a license (Apache 2.0, Gemma terms, or similar) that governs commercial reuse. What you generate and how you use it remains your responsibility, the same as any tool.

Do abliterated models get dumber? Slightly, and it varies by build. Abliteration removes a direction in activation space, and aggressive removal can nick adjacent capabilities. Well-made builds like the ones listed here measure the loss at a few benchmark points. Refusal-free finetunes like Dolphin sidestep the surgery but change model personality instead. Our Qwen 3.8 benchmarks breakdown covers what the unmodified base scores, which is your ceiling either way.

What is the minimum hardware to start? A 12GB GPU or a 16GB Apple Silicon Mac runs Hermes 4 14B or the Gemma 4 A4B build at usable speeds. The sweet spot in 2026 is 24GB of VRAM or 32GB unified memory, which unlocks the 24B to 27B class where entries #1, #2, and #4 live. You can also use Qwen 3.8 for free through hosted channels first to gauge whether the base model fits your work before downloading 17GB of weights.

Why not one of the older lists’ models, like WizardLM or Vicuna? Base model age. A 2023-era 13B finetune loses to a 2026 27B on every axis: reasoning, context length, coding, multilingual output. Uncensored builds inherit their base’s ceiling, so the ranking above starts from current bases and works down only where hardware demands it.

The bottom line

Qwen 3.8-27B Uncensored is the default answer in 2026: the newest base, real multimodal support, and quants that fit cards people own. Dolphin 3.0 is the finetune alternative with the deepest ecosystem, and Hermes 4 is the thinking operator’s pick, uncensored by philosophy and governable by system prompt. Below 16GB, take the Gemma 4 A4B build for speed or Mistral Small 3.2 abliterated for prose. And whichever you serve, remember it boots as an unguarded API on localhost: test it with Apidog like you would any endpoint you plan to trust.