Originally published on locallyuncensored.com
If you've looked at the Discover tab in any local-AI app and wondered why some Llama variants have abliterated in the name, this is the post that explains it. Plus the curated download list for 2026.
What Abliteration Actually Is
Modern instruction-tuned LLMs have a learned refusal direction in their residual stream. When a prompt activates that direction strongly enough, the model outputs "I cannot help with that." The direction was put there during RLHF.
Abliteration removes it via orthogonalisation. You take a corpus of refused prompts, isolate the activation pattern that distinguishes them from accepted prompts, then project that direction out of every weight matrix. The result is a model with the same training and capabilities but no longer prone to categorical refusal.
It's a clean technique - not a finetune, not a jailbreak, not a system-prompt trick. Original paper: "Refusal in Language Models Is Mediated by a Single Direction" (Arditi et al., 2024).
Abliterated vs Other Uncensored Approaches
| Method | How it works | Effort | Quality impact |
|---|---|---|---|
| Abliteration | Project out refusal direction | hours on GPU | 1-3% degradation |
| Full finetune (Dolphin, Hermes) | Re-train on uncensored corpus | days, expensive | Variable |
| LoRA finetune | Adapter on uncensored data | hours | Minor, reversible |
| Merge (Frankenmerges) | Combine multiple finetunes | hours | Highly variable |
| System prompt jailbreak | Persona-style instructions | None | Brittle |
Abliteration is the cleanest research-grounded option. Dolphin and Hermes are battle-tested production finetunes.
Recommended Abliterated Models (2026)
Qwen 3.6 Family
-
richardyoung/qwen3-14b-abliterated:q4_K_M - 9 GB, fits 12 GB VRAM, vision-capable. Comes in
:q4_K_M(chat) and:agent(tool-calling) tags via Ollama. - Qwen 3.6 27B Samantha (huihui-ai variant) - abliterated dense 27B with the Samantha personality finetune.
Gemma 4 Heretic
- Stabhappy/gemma-4-31B-it-heretic-Gguf - Gemma 4 31B base abliterated. ~17 GB at Q4_K_M. Native vision, tool calling.
- Gemma 4 26B MoE HERETIC - 26B brain with 4B active. Smaller VRAM peak, MoE-fast inference.
Llama 3.1 Family
-
mannix/llama3.1-8b-abliterated:q5_K_M - 5.7 GB. The most-pulled abliterated Llama on Ollama. Comes with
:agenttag for tool calling. - mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated - the canonical reference variant.
Hermes 3
Hermes 3 is technically a full finetune, not abliteration, but functions similarly:
- hermes3:8b via Ollama - 4.7 GB, fits 8 GB GPUs. Good chat default.
- hermes3:70b - 40 GB, needs 48 GB VRAM or aggressive quantisation.
GLM 5.1 Heretic
The newest entrant: huihui-ai/Huihui-GLM-5.1-abliterated-GGUF. The 754B MoE GLM 5.1 abliterated. 236 GB at IQ2_M - not consumer hardware, but if you have a Mac Studio M4 Ultra, it's the strongest open abliterated model period.
How to Download and Run
Path 1 - Ollama (one command)
ollama pull richardyoung/qwen3-14b-abliterated:q4_K_M
ollama run richardyoung/qwen3-14b-abliterated:q4_K_M
Path 2 - Locally Uncensored (one click)
Open Locally Uncensored, navigate to Model Manager > Discover > Text, click the UNCENSORED filter tab. The 34 curated abliterated GGUFs are all there with one-click download.
The new v2.4.0 Settings > Model Storage override lets you redirect the GGUF download folder if you want them on a separate drive.
Hardware Recommendations
| VRAM | Best Abliterated Pick | Why |
|---|---|---|
| 8 GB | Llama 3.1 8B abliterated Q4_K_M | Fits with headroom |
| 12 GB (RTX 3060) | Qwen 3 14B abliterated Q4_K_M | Sweet spot, ~15 tok/s |
| 16 GB | Gemma 4 31B Heretic Q4_K_M | Best general-purpose at this VRAM |
| 24 GB (RTX 3090/4090) | Gemma 4 31B Heretic Q5_K_M | Higher quality |
| 48 GB+ | Hermes 3 70B or GLM 5.1 Heretic IQ2 | Frontier-tier quality |
Common Questions
Will an abliterated model write me malware?
Probably not the way you're thinking. Abliteration removes the categorical refusal but the model still has training-time priors against obviously-bad outputs. The models work best for legitimate-but-edge-case use cases: security research, fiction with violence, medical questions the base model deflects, legal grey areas, adult creative writing.
Are abliterated models dangerous?
No more than the underlying base. Abliteration removes a layer of guardrails. The model's underlying knowledge is unchanged from the base.
Can I abliterate a model myself?
Yes. The technique is well-documented and the code is on GitHub (search abliterator). You need a GPU with the model loaded, a few thousand refused-vs-accepted prompt pairs, and a few hours.
Locally Uncensored is AGPL-3.0 licensed. Built by PurpleDoubleD. Bug reports on GitHub Discussions or in the Discord.
Top comments (0)