7 best open-weight LLMs for uncensored chats in 2026

Key takeaways

• Stock models from the big labs refuse often. Google's Gemma 4 31B scores 2.5 out of 10 for willingness on the UGI Leaderboard, and a decensored "heretic" build of it scores 10.

• On a 24GB graphics card, start with Gemma 4 31B or Qwen3.8-27B. Both are Apache 2.0, and their decensored builds come as Q4_K_M files under 19GB.

• Dolphin Mistral 24B Venice Edition was fine-tuned to take its rules from your system prompt, and its Q4_K_M file fits a 16GB card.

• DeepSeek V4 Flash is the most willing unmodified model here, at 7.2 out of 10, but its quantized files take 97GB to 155GB.

• Decensoring doesn't lift usage policies or the law. Meta's Muse Glimmer policy bars removing its safety measures, and sexual content involving minors or non-consenting real people is illegal in many places.

Which open-weight LLM is best for uncensored chats?

On a 24GB graphics card, start with a decensored build of Gemma 4 31B. It scores 10 out of 10 on the UGI Leaderboard's willingness test and runs from an 18.7GB file. Qwen3.8-27B's decensored build is the close alternative, with a higher creative-writing score.

"Uncensored" covers two kinds of model. Some are trained to follow the system prompt and rarely refuse. Others are mainstream models with refusals stripped out afterward, uploaded to Hugging Face under names like "heretic" and "abliterated."

How we ranked them

We ranked the seven on how freely each one answers (as released, or in its decensored build), its creative-writing standing, the hardware it needs, and its context length and license. The UGI Leaderboard, whose data was last updated October 17, 2026, scores willingness (W/10) as "How far a model can be pushed before it refuses to answer or deviates from instructions." It also gives a writing score. EQ-Bench Creative Writing v3, which calls itself "A LLM-judged creative writing benchmark," ranked 146 models by Elo as of October 29, 2026. It only scores models as released, so a decensored build's figure is its base model's.

File sizes are Q4_K_M GGUF downloads on Hugging Face unless noted. Meta's Muse Spark Open (816 billion parameters) and Mistral Large 4 (about 1 trillion) are too large for home hardware, so they aren't here.

The seven models at a glance

1. Gemma 4 31B (heretic): best overall on a 24GB card

Google released Gemma 4 on April 2, 2026, the first Gemma models under Apache 2.0. The stock 31B model refuses a lot, so download a decensored version such as coder3101's heretic build.

Strengths: That build scores 10 out of 10 for willingness on UGI, against 2.5 for Google's original, and its card reports 15 refusals on 100 test prompts, where the original refused 99. The Q4_K_M file is 18.7GB, and Google lists a 256K-token context window.

Limits: The base model's EQ-Bench Elo of 1,368.6 puts it 89th of 146, well behind Qwen3.8-27B and Muse Glimmer. On a 16GB card, you'd drop to the 15.3GB Q3_K_M file and keep the context short.

Best for: Role-play and fiction on one 24GB graphics card, with few refusals.

2. Qwen3.8-27B (heretic): best for writing quality at 27B

Alibaba's Qwen team published Qwen3.8-27B on Hugging Face in August 2026 under Apache 2.0. The stock model is cautious; llmfan46's heretic build strips most of that out.

Strengths: The base model's EQ-Bench Elo is 1,670.5, 42nd of 146 and about 300 points above Gemma 4 31B. The heretic card reports three refusals out of 100 test prompts, down from 91 for the original. Context is 262,144 tokens natively, extendable to one million with YaRN scaling. The Q4_K_M file is 17.6GB.

Limits: UGI hasn't scored Qwen3.8 or its decensored builds yet, so the refusal count comes from the uploader alone. Thinking mode is on by default and writes out reasoning before each reply, though Qwen's card says you can turn it off per request.

Best for: Long fiction and campaigns where the prose counts as much as the freedom.

3. Dolphin Mistral 24B Venice Edition: best for 16GB graphics cards

Dolphin is a long-running line of fine-tunes from dphn. The Venice Edition, built with Venice.ai on Mistral Small 24B, came out in June 2025, and its looser behavior comes from training.

Strengths: UGI gives it 7.8 for willingness. Its model card says the system prompt "is what you use to set the tone and alignment of the responses." The Q4_K_M file is 14.3GB, which fits a 16GB card, and a 10.7GB IQ3_M file fits 12GB.

Limits: Its UGI writing score of 32.2 trails the Gemma heretic build's 39.6, and EQ-Bench doesn't list it. dphn's newer Venice Role Play Uncensored, a role-play tune of this model, has a model card from May 2026 but no downloadable weights on Hugging Face as of late October.

Best for: Character chat on a 12GB or 16GB card, steered by a system prompt you write.

4. DeepSeek V4 Flash: best for big-memory machines

DeepSeek released V4 Flash in April 2026 under the MIT license. It's a mixture-of-experts model with 284 billion parameters, 13 billion of them active per token, and a context window of one million tokens.

Strengths: It's the most willing model here as released, at 7.2 out of 10 with reasoning switched off, and its UGI writing score of 54.6 is the top one in this list. It ranks 59th of 146 on EQ-Bench, with an Elo of 1,559.4.

Limits: Size rules out most home setups: Unsloth's GGUF files run from 96.8GB (Q2_K_XL) to 155.1GB (Q4_K_XL), so you need a Mac with 128GB or more of unified memory, or a PC with lots of system RAM to offload into. The October update, V4.1-Flash-1017, belongs to a larger V4.1 line and hasn't been scored on UGI.

Best for: Mac Studio owners and multi-GPU rigs that want stronger prose without a decensored build.

5. Hermes 4.3 36B: best for steering with a system prompt

Nous Research released Hermes 4.3 36B in November 2025. It's built on ByteDance's Seed-OSS-36B base model and licensed under Apache 2.0.

Strengths: With no abliteration, it reaches 6.2 for willingness on UGI. Nous reports that it answered 74.6% of questions on RefusalBench, its own test of prompts that other models commonly turn down, and lists "Much easier to steer and align" among the changes from Hermes 3.

Limits: Its UGI writing score of 31.4 sits below the Gemma heretic build's 39.6, and EQ-Bench hasn't scored this version. At 21.8GB, the Q4_K_M file is tight on a 24GB card once context is loaded, though a 17.6GB Q3_K_M file exists.

Best for: People who'd rather set the rules in a system prompt and keep an unmodified model.

6. Muse Glimmer 30B: best for prose quality out of the box

Meta hasn't published a new Llama on Hugging Face since Llama 4 in April 2025; its newer open models carry the Muse name. Muse Glimmer 30B, released in August 2026 under Apache 2.0, is distilled from Meta's larger Muse Spark.

Strengths: Its EQ-Bench Elo of 1,804.3 ranks 28th of 146, ahead of every other model in this list. Meta ships its own 4-bit GGUF, a 16.8GB file it targets at 24GB of VRAM.

Limits: It refuses more than anything else here, at 4.2 out of 10 for willingness. Decensored uploads exist, but Meta's usage policy bars using the model to "intentionally circumvent or remove usage restrictions or other safety measures." The same policy says it "is not intended for individuals under the age of 18."

Best for: Fiction that stays inside mainstream limits, where the writing comes first.

7. gpt-oss-20b (heretic): best for laptops and 16GB machines

OpenAI released gpt-oss-20b in August 2025 under Apache 2.0. It's a mixture-of-experts model with 21 billion parameters and 3.6 billion active, and OpenAI says it runs "within 16GB of memory."

Strengths: The original scores 0.5 to 1.8 for willingness, depending on reasoning effort. p-e-w's heretic build scores 9.5, and its card reports three refusals out of 100 against 98 for the original. A 12.1GB MXFP4 GGUF of that build is available.

Limits: Creative writing is its weak spot. The base model's EQ-Bench Elo is 664.5, 143rd of 146, and the heretic build's UGI writing score is 24.1. OpenAI's usage policy for the model is short: "By using OpenAI gpt-oss-20b, you agree to comply with all applicable law."

Best for: Short chats on a laptop or a 16GB card when nothing larger fits.

Abliteration and fine-tuning, in plain terms

Abliteration removes a model's refusals without retraining it. Maxime Labonne's explainer on Hugging Face walks through the method: run the model on prompts it refuses and prompts it answers, find the internal "refusal direction" that separates the two, then edit the weights so the model can't express that direction anymore. In his example, abliteration "degraded the model's quality," and extra training recovered most of it.

Heretic, the tool behind the "heretic" builds above, automates abliteration. Its README says it picks settings by "co-minimizing the number of refusals and the KL divergence from the original model," KL divergence being a measure of how far the edited model's answers drift from the original's on ordinary prompts.

Fine-tuning is the other route: Dolphin and Hermes are trained further so they follow the system prompt by design. Community role-play tunes work the same way, and TheDrummer's Behemoth-X 123B v2 scores 8.2 for willingness on UGI.

How to run these models locally

All seven come as GGUF files, the format llama.cpp and the apps built on it read.

• Ollama: An app based on llama.cpp that runs a Hugging Face GGUF with `ollama run hf.co/{username}/{repository}:{quantization}`, defaulting to Q4_K_M when a repository has it.

• LM Studio: A desktop app with a model browser and chat window, powered by "MLX and llama.cpp under the hood" (MLX is Apple's framework for Apple Silicon).

• llama.cpp: The engine underneath, with an `-hf` option that downloads models from Hugging Face and an OpenAI-compatible server mode.

• KoboldCpp: A single executable "built on llama.cpp" that bundles the KoboldAI Lite interface, with memory, world info, and characters for stories and role-play.

• SillyTavern: A front end that doesn't run models; its docs call it "only an interface." It connects to KoboldCpp, llama.cpp, Ollama, and other backends, and it's built around character cards that set a persona's behavior.

The file sizes above are a floor, since context needs memory too. Our VRAM guide lists what each graphics card holds, and the Mac vs. PC guide covers how unified memory lets a 128GB Mac load models no consumer card can. Two or three quantizations of a 20GB model add up fast; see how much storage local AI models need.

Licenses, usage policies, and the law

Six of the seven are Apache 2.0, and DeepSeek V4 Flash is MIT. Neither license sets content rules. Apache 2.0 says "You are solely responsible for determining the appropriateness of using or redistributing the Work." Meta and OpenAI add usage policies on top, quoted in their entries above.

A decensored model doesn't change what's legal. Sexual content involving minors, and sexual content depicting real people without their consent, is illegal in many places no matter what a model allows.

Open-weight LLMs for uncensored chats FAQ

What is the best uncensored open-weight LLM?

On a 24GB card, a heretic build of Gemma 4 31B, which scores 10 out of 10 for willingness on the UGI Leaderboard. Qwen3.8-27B's heretic build is the alternative if writing quality comes first, and Dolphin Mistral 24B Venice Edition fits a 16GB card.

Does abliteration make a model worse?

It can, a little. Labonne reported a drop in quality in his original example. Heretic builds publish a KL divergence to show the drift; coder3101's Gemma 4 31B build lists 0.0434.

How much VRAM do I need to run a 30B model locally?

About 24GB for a Q4_K_M file of a 27B to 31B model with room for context. Gemma 4 31B's file is 18.7GB and Qwen3.8-27B's is 17.6GB. On a 16GB card, use a Q3 file with a short context, or a 24B model like Dolphin at Q4_K_M.

Can I use these models commercially?

Apache 2.0 and MIT both allow it, and that covers all seven. Muse Glimmer and gpt-oss-20b also carry usage policies.