The best uncensored LLMs for your hardware.

Quick picks

How to read the sizes below. "File" is the size of the quantized download, as published by its maintainer or a cited guide. "Memory" is a YouRunAI estimate: the file plus room for a few thousand tokens of context and runtime overhead. Long contexts need more. We haven't physically tested these models on our own hardware yet. When we do, this page will label each result Tested, with the date.

8GB VRAM (or a 16GB Mac)

Dolphin 3.0 Llama 3.1 8B: the easiest start

• Method: filtered fine-tune by Eric Hartford and Cognitive Computations. The model card says the person running it sets the system prompt and the alignment.
• Get it: ollama run dolphin3. The Ollama library build is Q4_K_M, a 4.9GB file with a 128K context window. Q8_0 (8.5GB) and FP16 (16GB) tags exist too.
• Memory: about 6 to 7GB at a modest context (estimated).
• License: Llama 3.1 Community License.
• Best for: general chat, writing, and learning how a system prompt shapes an uncensored model.

Qwen3-8B abliterated v2: a stronger reasoner at the same size

• Method: abliteration by huihui-ai.
• Repo: huihui-ai/Huihui-Qwen3-8B-abliterated-v2. InsiderLLM reports about 5GB at Q4_K_M.
• Memory: about 6 to 7GB (estimated).
• Best for: Q&A and step-by-step reasoning.

12GB VRAM

Gemma 4 12B abliterated

• Method: abliteration by huihui-ai.
• Repo: huihui-ai/Huihui-gemma-4-12B-it-abliterated, with GGUF quantizations from mradermacher. InsiderLLM reports that Q4 and Q5 fit a 12GB card comfortably.
• Best for: a capable all-rounder that leaves room for context on a 12GB card.

Qwen3-8B abliterated v2 at Q8_0

The same model as above at near-full precision. The file is about 9GB (InsiderLLM). Pick this one if quality matters more to you than context length.

16GB VRAM (or a 24GB Mac)

Dolphin 3.0 Mistral 24B

• Method: filtered fine-tune.
• Repo: dphn/Dolphin3.0-Mistral-24B, released under Apache 2.0. There's a reasoning variant, Dolphin3.0-R1-Mistral-24B, trained on DeepSeek-R1 reasoning traces.
• File: Q4 quantizations run about 14 to 15GB (Apidog). Use a smaller quant on a 16GB card.
• Tip: the maintainers recommend a very low temperature, 0.05 to 0.1, for Mistral-24B-based Dolphin models.
• Best for: general chat, coding help and agent work, with tool calling supported.

Gemma 4 26B-A4B heretic

• Method: Heretic, an automated refinement of abliteration.
• Repo: mradermacher/gemma-4-26B-A4B-it-heretic-GGUF. huihui-ai also publishes abliterated Gemma 4 builds.
• Why it's fast: it's a mixture-of-experts model with 26B total parameters but only about 4B active per token. Apidog puts it at 12 to 16GB.
• Trade-off: it trails the dense 24B to 27B models on hard reasoning.

24GB VRAM (RTX 3090, 4090, 5090) or a 32GB Mac

Qwen 3.8-27B uncensored: the current leader

Alibaba released Qwen 3.8-27B's weights on August 14, 2026. Abliterated, quantized builds appeared within hours. Apidog calls it the first time the strongest refusal-free local model is close to the strongest open model, full stop.

• Method: abliteration.
• Repos: orcarouter/Qwen3.8-27B-Uncensored-GGUF, plus a more aggressive variant, 0xKitkat/Qwen3.8-27B-Uncensored-Aggressive. The base model is Qwen/Qwen3.8-27B.
• File (reported by Apidog): Q4_K_M is 16.8GB (good for 24GB cards or 32GB Macs). IQ4_XS is 15.3GB (a 16GB card, tight). Q3_K_M is 13.5GB.
• Watch out: it needs a llama.cpp build from May 2026 or later. Older LM Studio or Ollama versions may fail to load it, so update first.
• Best for: the strongest local all-rounder, covering coding, agents and long reasoning.

Qwen 3.6-27B abliterated: the proven fallback

• Method: abliteration by huihui-ai.
• Get it: ollama run huihui_ai/qwen3.6-abliterated:27b, as given on the model card. Or use GGUF from huihui-ai/Huihui-Qwen3.6-27B-abliterated-MTP-GGUF. InsiderLLM reports Q6_K at about 22.4GB.
• Best for: a widely used daily driver, and the safe choice if your runtime can't load Qwen 3.8 yet.

48GB and up (two GPUs, pro cards, big Macs)

• Hermes 4 70B (NousResearch/Hermes-4-70B). This isn't abliterated. It's trained with neutral alignment and follows your system prompt instead of a vendor policy. The 14B version fits a 12GB card at Q4 (Apidog).
• DeepSeek-R1-Distill-Llama 70B abliterated (bartowski/huihui-ai_DeepSeek-R1-Distill-Llama-70B-abliterated-GGUF). InsiderLLM reports it runs at Q4 on two RTX 3090s.
• Server-scale models like DeepSeek V4.1 Flash uncensored and GLM-5.3 UNCENSORED need multi-GPU servers or an API. They're beyond a desktop, so we cover them on their own pages.

How we choose

We pick models that meet all of these:

1. Real, downloadable weights with a named maintainer and a working run path (Ollama, LM Studio or llama.cpp).
2. A clearly documented method: fine-tune, abliteration, heretic, or neutral alignment.
3. Several independent sources agree on quality and size. We link the sources for every number above.
4. Active maintenance: a range of quantizations and a model card that's actually filled in.

Every model here can produce harmful, false or offensive output. They have no default safety layer, so you are the safety layer. Read Is uncensored AI legal and safe? before you deploy one for other people.

Next step

Pick a model above and follow Run uncensored models with Ollama. It's one install and one command.

Sources

• Apidog: the 7 best uncensored LLMs you can run locally in 2026 (August 19, 2026)
• InsiderLLM: best uncensored local LLMs by VRAM tier (updated July 18, 2026)
• Ollama: dolphin3 tags
• Dolphin 3.0 Mistral 24B on Featherless, a summary of the model card, including the temperature advice
• Huihui Qwen3.6-27B abliterated model card, for the Ollama tag and usage warnings