Best uncensored LLM.
Verdict (September 2026): on a 24GB GPU, Qwen 3.8 27B Obliterated is our top pick (~16 GB at Q4). On 12GB cards, Mistral Nemo 12B is permissively aligned out of the box and Apache 2.0. Below: how "uncensored" actually works, the model families that matter, and the part nobody puts in the title — what stays your responsibility.
What "uncensored" really means
Two techniques dominate. Refusal-stripped fine-tuning retrains a base model on datasets with refusals removed — the approach popularized by the Dolphin series. Abliteration is newer and cheaper: it identifies the internal "refusal direction" in a model’s activations and removes it from the weights directly, no retraining. Neither adds capability; both lower the model’s reflex to decline a prompt. Benchmarks typically come out flat to slightly lower than the base model.
The picks, by VRAM tier
| Tier | Pick | Why |
|---|---|---|
| 24GB (RTX 3090/4090) | Qwen 3.8 27B Obliterated | Full 27B-class capability, ~16 GB at Q4, refusal direction removed |
| 12GB (RTX 3060/5070) | Mistral Nemo 12B | Permissive alignment by default, Apache 2.0, ~8 GB at Q4 — no surgery needed |
| 8–12GB, reasoning-first | Qwythos 9B abliterated (GGUF) | Qwen 3.5 finetune distilled from Claude reasoning traces — ~5 GB at Q4, 1M context, refusal direction removed |
| Community fine-tunes | Dolphin family (Cognitive Computations) | The reference refusal-stripped series, built on Mistral/Mixtral/Llama bases |
| Neutral alignment | Nous Hermes family (Nous Research) | Tuned for instruction-following with minimal moralizing, strong all-rounders |
| DIY route | Abliterated variants of Qwen/Llama/Gemma on Hugging Face | Search "abliterated" + your base model; check downloads and community feedback first |
Not sure what your GPU holds? Run the numbers in the LLM VRAM calculator, or start from your hardware with the 8GB VRAM and RTX 3090 rankings.
Who actually needs this
Fiction and roleplay writers tired of a model refusing a villain’s dialogue; security researchers and red-teamers who need a model that will discuss attack techniques they are paid to defend against; and anyone whose legitimate prompts trip overcautious filters. If your use case is none of these, a standard model plus a good system prompt usually gets you there with less quality loss.
Your responsibility, stated plainly
Everything an uncensored model outputs on your machine is on you. The license still applies, your local laws still apply, and “the model let me” is not a defense anywhere. We list these models for the legitimate uses above — creative work, research, red-teaming — and running them locally is precisely what keeps that use private and under your control.
Frequently asked questions
What does an "uncensored" LLM actually mean?
A model tuned to refuse less. Two main routes exist: fine-tuning on data with refusals stripped out (the Dolphin approach), and "abliteration" — surgically removing the internal refusal direction from an existing model's weights without retraining it. Neither adds knowledge; both remove the reflex to decline.
What is the best uncensored LLM for 24GB of VRAM?
Qwen 3.8 27B Obliterated is our current pick for a 24GB GPU (RTX 3090/4090 class): roughly 16 GB at Q4 quantization, with the underlying model's full general capability. On 12GB cards, Mistral Nemo 12B is the pragmatic choice — permissively aligned out of the box, Apache 2.0, about 8 GB at Q4.
Are uncensored models legal to use?
Running an open-weight model locally is legal; what you generate with it remains your responsibility under your local laws and the model's license. "Uncensored" removes the model's reflex to refuse — it does not remove your obligations.
Do uncensored models perform better than the originals?
On benchmarks they usually score the same or slightly lower than their base model — abliteration can nick reasoning quality. People run them for fewer spurious refusals in fiction, roleplay, security research and red-teaming, not for extra capability.
By Mohamed Meguedmi — independent comparator of locally-runnable LLMs, benchmarked on a real RTX 5070 Ti (data CC BY 4.0). See the local LLM leaderboard.