The highest-scoring uncensored model you can run at home is not a 70-billion-parameter monster that needs two graphics cards. It is a 31B edit of Google's Gemma 4 that fits on one 24 GB card. On the UGI leaderboard, the community benchmark people usually mean when they search for an uncensored LLM leaderboard, gemma-4-31B-it-heretic scores 65.7. That puts it fifth out of 1,027 models, above every 70B model on the board and half a point behind the best 123B one.
So finding the best uncensored LLM to run locally is mostly a question of what your machine can hold, and past a certain point extra memory stops helping. This guide sorts the tested models by memory size, from an 8 GB laptop to a 128 GB workstation. The scores come from the UGI leaderboard's data file as it stood on November 10, 2026. Download sizes come from Hugging Face and Ollama, checked the same day.
Three scores do most of the work
UGI stands for Uncensored General Intelligence. One maintainer, who goes by DontPlanToEnd, runs it as a Hugging Face Space, and the test questions are kept private so that nobody can train a model on them. The data file lists 1,329 tested settings across 1,027 models. Three columns matter for this guide:
- W/10, willingness. How far a model can be pushed before it refuses or stops following the instructions. The scale tops out at 10.
- NatInt, natural intelligence. General knowledge and reasoning: textbook facts, pop culture, and a set of real-world estimation tasks.
- UGI. The headline number. It combines willingness with how much the model knows about sensitive subjects.
A model can know a great deal and still refuse to use it, so the picks below look at willingness first and intelligence second. Every score quoted is a model's best tested setting.
Only 26 models score a perfect 10 for willingness. All 26 are community edits of open-weight models. None was released that way by a lab.
The best uncensored local LLM for each memory size
The rule for local models is simple: the model file has to fit in your graphics card's memory, or in a Mac's unified memory, with a few gigabytes to spare for the conversation itself. The sizes below are for Q4_K_M, a 4-bit compressed format and the one Ollama picks by default when you pull a model from Hugging Face.
| Your memory | Size that fits | Best tested pick | W/10 | NatInt | Download |
|---|---|---|---|---|---|
| 8 GB laptop | 4B | Qwen3.5-4B-SOMPOA-heresy-v2 | 10 | 22.4 | about 3.3 GB |
| 8 GB card, 16 GB Mac | 9B | Qwen3.5-9B-heretic-v2 | 9.5 | 24.2 | 5.6 GB |
| 16 GB card | 24B | WeirdCompound-v1.7-24b | 8 | 30.2 | about 15 GB |
| 24 GB card, 32 GB Mac | 27B to 31B | gemma-4-31B-it-heretic | 10 | 36.5 | 18.7 GB |
| Two 24 GB cards, 64 GB Mac | 70B | Assistant_Pepe_70B | 9.5 | 35.3 | about 43 GB |
Each pick is the smartest model in its size class that scores 9 or more on willingness, with the 4B a near tie decided by its perfect 10. The 16 GB row breaks the rule, for a reason explained below. Sizes marked "about" are Ollama's listing for the same base model family at the same compression. The others are the actual files on Hugging Face.
Read down the NatInt column and the pattern is plain. Intelligence climbs up to the 24 GB row, then stops.
8 GB: small Qwen builds that almost never say no
For a laptop with 8 GB of memory, Qwen3.5-4B-SOMPOA-heresy-v2 is the standout. It scores a perfect 10 for willingness, and its model card reports 4 refusals on a 104-prompt test set, against 103 for the original Qwen3.5 4B. The edit cost little: official Qwen3.5 4B scores 23.9 on NatInt with thinking on, and this version scores 22.4.
With an 8 GB graphics card or a 16 GB Mac you can step up to Qwen3.5-9B-heretic-v2. Its card reports 6 refusals out of 100 where the original refused all 100, and UGI gives it 9.5 for willingness. The price was higher this time. Official Qwen3.5 9B scores 31 on NatInt and the edited one 24.2, which leaves it only slightly ahead of the 4B. If the 9B is a squeeze on your machine, you lose very little by staying with the smaller model.
16 GB: the awkward middle
Models of about 24B, mostly merges and fine-tunes from the Mistral Small family, own this bracket. None of them combines top willingness with real brains: among the 24B models that score 9 or more on willingness, the best NatInt is 26.4. The better compromise is WeirdCompound-v1.7-24b, a merge of storytelling models, at 8 for willingness and 30.2 for NatInt. Its writing score of 47 is the highest of any model at 24B or below, so it suits fiction more than research.
A 24B model at Q4_K_M is about a 15 GB download, which leaves a 16 GB card almost no room. Most people drop to a slightly smaller quantization.
The model to watch here is Gemma 4 12B, which arrived in June 2026 and is an 8 GB download at Q4_K_M. The official version scores 29.5 on NatInt and just 2.2 on willingness. Uncensored builds are easy to find, and one GGUF repo shows about 98,000 downloads in the past month, but UGI has not tested any of them. It is probably the right answer for 12 to 16 GB. Nobody has measured it yet.
24 GB: where the leaderboard peaks
gemma-4-31B-it-heretic, uploaded by coder3101 within a day of Gemma 4's release on April 2, 2026, is the model everything else on this page gets compared with.
- Willingness: 10 out of 10. Official Gemma 4 31B scores 1.2.
- Intelligence: 36.5 on NatInt, against 38.9 for the official model. Removing the refusals cost 2.4 points.
- Refusals: the model card reports 15 refusals on 100 test prompts, down from 99.
- Size: 18.7 GB at Q4_K_M, which leaves roughly 5 GB free on a 24 GB card.
- Licence: Apache 2.0, the same as Gemma 4 itself.
A near-identical build from llmfan46 scores 65.0, and its GGUF repo shows about 60,600 downloads in the past month.
One setting matters more than any other: leave thinking switched on. With reasoning enabled the model scores 65.7 on UGI. With it off, the same weights score 56.4.
Two alternatives fit the same card:
Qwen3.5-27B-hereticscores 10 for willingness and 35.5 on NatInt. At 16.5 GB it leaves more room for long conversations. Its card reports 14 refusals out of 100, down from 94.gemma-4-26B-A4B-it-ultra-uncensored-hereticscores 10 for willingness and 35.7 on NatInt. Its writing score is 45.9 against 40.5 for the 31B, so it is the better pick for stories. According to Google's model card, only 3.8B of its 25.2B parameters are active for each token, the design that makes this kind of model quicker than a dense one of the same size.
How smart is the best willing model at each size?
Tap a row for details.
- Willingness 10 and a UGI score of 65.7, the best of anything that fits one 24 GB card.
UGI leaderboard NatInt score, best tested setting per model. Data file as of November 10, 2026.
Past 24 GB, extra memory buys very little
A 70B model is about a 43 GB download, so it needs two 24 GB cards or a 64 GB Mac. The best one on the board, Assistant_Pepe_70B, a fine-tune of Llama 3.1 70B, scores 59.5 on UGI with a NatInt of 35.3. The highest NatInt of any 70B model is 37.9. Twice the memory gets you a lower uncensored score than the 31B Gemma and about the same intelligence.
At roughly 120B the picture barely changes. The most intelligent build there with willingness of 9 or more, a derestricted version of gpt-oss-120b, has a NatInt of 35.7. Mistral Large 2411 deserves a mention because it scores 7.5 on willingness with no modification at all, and 36.2 on NatInt. It is a 73 GB download, and its licence allows non-commercial use only.
You have to reach server hardware before the numbers move:
- Qwen3.5-397B-A17B-heretic is the smartest fully willing model on the board, with willingness 10 and NatInt 39.5. That is 3 points above the 31B Gemma for a model nearly 13 times the size.
- GLM-4.6-Derestricted-v3, at 355B, has the highest UGI score of any downloadable model: 69.8.
- DeepSeek V4 Flash, at 284B, scores 7.2 on willingness as released, with reasoning off, and 47.9 on NatInt. No open-weight model with willingness of 7 or more scores higher on NatInt, and nobody is running it on a gaming PC.
The likely reason is supply. Community editors can only work with the open models that exist, and the open models clearly smarter than Gemma 4 31B have hundreds of billions of parameters. There is not much in between.
Smart or uncensored: no local model gets both
The smartest model on the board is GPT-6 Astra, with a NatInt of 81.9. Its willingness score is 2.8. Gemini 4 Argon is close behind at 81.4, and scores 1.8.
The best local uncensored pick scores 36.5. That is less than half. Across all 1,027 models, not one scores 7 or more on willingness and 55 or more on NatInt. The closest is a closed model, Grok 4.20 with reasoning, at 7.5 and 52.8.
Read plainly, the scores say this. A local uncensored model will answer the question a big-lab model lectures you about, write the scene it would refuse, and keep your prompts on your own machine. It will also know less and reason worse than the model that refused you. Which of those matters more depends on the job.
What the leaderboard hasn't tested yet
UGI is a one-person project and its pace has slowed. The data file shows 12 settings tested in October 2026 and 4 in August. Since July, the only locally runnable model added is Meta's Muse Glimmer 30B. So this ranking is a snapshot of the models of spring 2026, and some popular newer ones have no score:
- Qwen 3.8 27B. The official repo went up on Hugging Face on August 5, 2026, under Apache 2.0. Three uncensored builds of it each show more than a million downloads in the past month, far more than any tested pick above.
- Gemma 4 12B, covered in the 16 GB section.
- Muse Glimmer 30B. As released it scores 4.2 on willingness and 37.2 on NatInt, so it is more willing out of the box than Gemma 4 31B and about as smart. Abliterated builds exist, one with about 41,900 downloads in the past month. None has been tested.
Popular and tested are different things. For a build with no score, the model card is the next best evidence.
Most of these models are made with Heretic, an open-source tool that automates abliteration. It finds the internal direction a model uses to refuse and removes it from the weights, then searches for the settings that cut refusals while changing the model as little as possible. Its README says a 4B model takes 30 to 40 minutes on an RTX 3090.
Cards made with Heretic report two numbers. Refusals, out of about 100 test prompts, should be low. KL divergence, which measures how far the edited model's answers drift from the original's, should be close to zero. The 31B Gemma pick reports 0.0434. Our explainer on what an abliterated model is walks through the method.
From download to first answer
Run your first uncensored local model
Five steps, most of them waiting for a download.
Look up your graphics card's memory, or your Mac's total memory. That number decides which row of the table you shop from.
With Ollama installed, one command pulls the top pick from Hugging Face and starts a chat:
bashollama run hf.co/mradermacher/gemma-4-31B-it-heretic-GGUF:Q4_K_M
Swap in another repo name and quantization tag for the other models on this page. LM Studio does the same job with a search box and a download button. It needs an Apple Silicon Mac on macOS 14 or newer, or a Windows PC whose processor supports AVX2, and it recommends 16 GB of RAM.
No GPU at all
Disclosure: Cleus is our product.
If none of these fits your hardware, the other route is a hosted uncensored assistant. Cleus is one. It runs in the browser, has a free plan, and there is nothing to download. It has no row on the UGI leaderboard, so there is no score to quote here and we won't guess one. What you can do is put the same prompt to Cleus and to a big-lab model, side by side, on our uncensored AI leaderboard page.
Be clear about the trade. A local model is the only setup where no company can see your prompts. With any hosted service, Cleus included, your chats sit on someone else's servers unless you turn on end-to-end encryption, which Cleus offers on Premium. Hosted services also block illegal content. Our roundup of the best uncensored AI chatbots compares the hosted options.
Questions about running an uncensored LLM locally
What is the best uncensored LLM to run locally in 2026?
On tested scores, gemma-4-31B-it-heretic. It has a perfect willingness score, the highest UGI score of any model that fits on one consumer graphics card, and an 18.7 GB download. Newer builds based on Qwen 3.8 27B are far more popular and may be better, but the UGI leaderboard has not scored them.
What is the best uncensored model for 8 GB of VRAM?
Qwen3.5-9B-heretic-v2 at 5.6 GB, or the 4B Qwen3.5-4B-SOMPOA-heresy-v2 if you want room to spare. The two are close on intelligence: 24.2 against 22.4.
What is the UGI leaderboard?
An independent benchmark hosted on Hugging Face and run by DontPlanToEnd. It scores models on willingness, knowledge of sensitive topics, general intelligence and writing, using private questions. Models are added by hand, so new releases can wait months for a score.
Does abliteration make a model less intelligent?
Usually a little, sometimes a lot. In UGI's data the 27B Qwen lost 0.3 NatInt points and the 31B Gemma lost 2.4, while the 9B Qwen lost almost 7. Results vary by model and by who made the edit, which is why tested builds are worth preferring.
Are uncensored models on Hugging Face safe to download?
Anyone can upload one. Stick to uploaders with a long record, check that the model card reports refusal and KL divergence figures, and keep Ollama or LM Studio up to date. Hugging Face's own Ollama guide points to established community quantizers such as bartowski.
Is it legal to run an uncensored AI model?
In most places, yes. Downloading and running an open-weight model is no different from running any other software, and what you do with its answers is covered by the same laws as anything else. Check the licence if you plan to use a model for work: Mistral Large 2411, for one, is non-commercial only.
Would you rather not manage any of this? Cleus is an uncensored AI that runs in your browser.
Sources
- UGI Leaderboard (Uncensored General Intelligence) by DontPlanToEnd
- coder3101/gemma-4-31B-it-heretic model card
- mradermacher/gemma-4-31B-it-heretic-GGUF files
- llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF
- google/gemma-4-31B-it model card
- llmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic model card
- Bobi099/Qwen3.5-27B-heretic model card
- mradermacher/Qwen3.5-27B-heretic-GGUF files
- trohrbaugh/Qwen3.5-9B-heretic-v2 model card
- filvyb/Qwen3.5-9B-heretic-v2-GGUF files
- MuXodious/Qwen3.5-4B-SOMPOA-heresy-v2 model card
- FlareRebellion/WeirdCompound-v1.7-24b model card
- SicariusSicariiStuff/Assistant_Pepe_70B model card
- mistralai/Mistral-Large-Instruct-2411 model card
- Qwen/Qwen3.8-27B on Hugging Face
- Heretic, the abliteration tool, on GitHub
- Hugging Face docs: use Ollama with any GGUF model
- Ollama library: Gemma 4 sizes
- Ollama library: Qwen3.5 sizes
- Ollama library: Llama 3.3 70B size
- Ollama library: Mistral Large size
- Ollama library: Mistral Small 3.2 size
- LM Studio system requirements
- Most uncensored AI models, ranked and explained, on Cleus
Ask the question this article didn’t answer.
Ask anything about best uncensored llm. Cleus opens this article, searches the live web and gives you a straight answer, with sources.
