DefiledAI
UNCENSORED MODEL DATABASE
27 models — abliterated, Dolphin, and uncensored fine-tuned open-weight models. Community-tested with quality retention scores and direct HuggingFace links.
What is Abliteration?
Representation engineering that removes refusal direction vectors from model weights. No training required — takes minutes. Quality retention is typically 96–99% of the base model. Technical explainer →
Tags
27 models found
NEW
Llama 3.1 70B Abliterated
ABLITERATIONby failspy · Llama Abliterated
★★★★★4.9
412 votes
PARAMS
70B
MIN VRAM
40.2GB (Q4_K_M)
QUALITY RETENTION
98%
QUANTS
Q2_KQ4_K_MQ5_K_MIQ3_M
Best overall abliterated model. v3 improves reasoning retention.
Dolphin 2.9 Llama 3.1 8B
FINE-TUNE (DOLPHIN)by cognitivecomputations · Dolphin
★★★★★4.8
334 votes
PARAMS
8B
MIN VRAM
5.5GB (Q4_K_M)
QUALITY RETENTION
98%
QUANTS
Q4_K_MQ5_K_MQ8_0
Most popular uncensored 8B. Best entry-level Dolphin model.
Llama 3.1 8B Abliterated
ABLITERATIONby failspy · Llama Abliterated
★★★★★4.8
312 votes
PARAMS
8B
MIN VRAM
5.5GB (Q4_K_M)
QUALITY RETENTION
99%
QUANTS
Q4_K_MQ5_K_MQ8_0F16
FailSpy abliteration. 99.2% MMLU retention. Community favourite.
Mistral 7B v0.3 Abliterated
ABLITERATIONby failspy · Mistral Abliterated
★★★★★4.8
287 votes
PARAMS
7B
MIN VRAM
4.8GB (Q4_K_M)
QUALITY RETENTION
99%
QUANTS
Q4_K_MQ5_K_MQ8_0F16
Cleanest abliteration at 7B. 99.2% MMLU retention.
DeepSeek R1 70B Abliterated
ABLITERATIONby community · DeepSeek Abliterated
★★★★★4.8
241 votes
PARAMS
70B
MIN VRAM
40GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ5_K_M
Best uncensored reasoning model locally. Needs 48GB VRAM.
Dolphin 2.9 Llama 3.1 70B
FINE-TUNE (DOLPHIN)by cognitivecomputations · Dolphin
★★★★★4.7
198 votes
PARAMS
70B
MIN VRAM
40.2GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ5_K_M
Best quality Dolphin. Needs 48GB VRAM (dual 3090 NVLink).
Dolphin 2.8 Mistral 7B
FINE-TUNE (DOLPHIN)by cognitivecomputations · Dolphin
★★★★★4.7
287 votes
PARAMS
7B
MIN VRAM
4.8GB (Q4_K_M)
QUALITY RETENTION
98%
QUANTS
Q4_K_MQ5_K_MQ8_0
Classic Dolphin on Mistral. Fast, reliable, widely tested.
NEW
Qwen 3 32B Abliterated
ABLITERATIONby community · Qwen Uncensored
★★★★★4.7
115 votes
PARAMS
32B
MIN VRAM
21GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ5_K_M
Top coding model that fits 24GB single GPU. Q4 = 21GB.
Qwen 2.5 72B Uncensored
ABLITERATIONby community · Qwen Uncensored
★★★★★4.7
198 votes
PARAMS
72B
MIN VRAM
41GB (Q4_K_M)
QUALITY RETENTION
98%
QUANTS
Q4_K_MQ5_K_M
Best uncensored coding model at 72B scale.
NEW
Dolphin 3.0 Llama 3.2 3B
FINE-TUNE (DOLPHIN)by cognitivecomputations · Dolphin
★★★★★4.6
89 votes
PARAMS
3B
MIN VRAM
2.2GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ5_K_MQ8_0
Newest Dolphin on Llama 3.2 3B. Runs on 4GB VRAM. Very fast.
NEW
Qwen 3 8B Abliterated
ABLITERATIONby community · Qwen Uncensored
★★★★★4.6
108 votes
PARAMS
8B
MIN VRAM
5.5GB (Q4_K_M)
QUALITY RETENTION
98%
QUANTS
Q4_K_MQ5_K_MQ8_0
Qwen 3 8B abliterated. Thinking mode available.
NEW
Qwen 3 14B Abliterated
ABLITERATIONby community · Qwen Uncensored
★★★★★4.6
87 votes
PARAMS
14B
MIN VRAM
9.5GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ5_K_M
Best mid-range Qwen 3 abliteration. Fits single 12GB GPU.
DeepSeek R1 7B Abliterated
ABLITERATIONby community · DeepSeek Abliterated
★★★★★4.6
167 votes
PARAMS
7B
MIN VRAM
5GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ8_0
Uncensored reasoning at 7B. Chain-of-thought fully intact.
Dolphin 2.9 Mixtral 8x22B
FINE-TUNE (DOLPHIN)by cognitivecomputations · Dolphin
★★★★★4.5
112 votes
PARAMS
141B
MIN VRAM
43GB (Q4_K_M)
QUALITY RETENTION
96%
QUANTS
Q2_KQ4_K_M
Dolphin on MoE. Exceptional quality when hardware allows.
Llama 3.2 3B Abliterated
ABLITERATIONby bartowski · Llama Abliterated
★★★★★4.5
134 votes
PARAMS
3B
MIN VRAM
2.2GB (Q4_K_M)
QUALITY RETENTION
99%
QUANTS
Q4_K_MQ5_K_MQ8_0
Best tiny abliterated model. Runs on 4GB VRAM at very high speed.
Mistral Nemo 12B Abliterated
ABLITERATIONby bartowski · Mistral Abliterated
★★★★★4.5
134 votes
PARAMS
12B
MIN VRAM
8.1GB (Q4_K_M)
QUALITY RETENTION
98%
QUANTS
Q4_K_MQ5_K_MQ6_K
128K context retained. Strong mid-range abliterated option.
NEW
Qwen 3 4B Abliterated
ABLITERATIONby community · Qwen Uncensored
★★★★★4.5
92 votes
PARAMS
4B
MIN VRAM
3GB (Q4_K_M)
QUALITY RETENTION
98%
QUANTS
Q4_K_MQ5_K_MQ8_0
Qwen 3 4B with refusal removed. Strong coding at tiny size.
Gemma 3 12B Abliterated
ABLITERATIONby community · Gemma Abliterated
★★★★★4.5
91 votes
PARAMS
12B
MIN VRAM
8.1GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ5_K_MQ8_0
Gemma 3 12B abliterated. Strong reasoning for size.
WizardLM-2 8x22B
FINE-TUNEby microsoft · WizardLM
★★★★★4.5
98 votes
PARAMS
141B
MIN VRAM
43GB (Q4_K_M)
QUALITY RETENTION
96%
QUANTS
Q2_KQ4_K_M
MoE architecture. Exceptional capability when hardware allows.
Mistral Small 3.1 Abliterated
ABLITERATIONby community · Mistral Abliterated
★★★★☆4.4
78 votes
PARAMS
24B
MIN VRAM
16GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ5_K_M
Mistral's newest small model with abliteration. 128K context.
Gemma 3 4B Abliterated
ABLITERATIONby community · Gemma Abliterated
★★★★☆4.4
78 votes
PARAMS
4B
MIN VRAM
3GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ5_K_MQ8_0
Google's newest Gemma 3 abliterated. 4GB VRAM.
Gemma 2 27B Abliterated
ABLITERATIONby community · Gemma Abliterated
★★★★☆4.4
112 votes
PARAMS
27B
MIN VRAM
18GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ5_K_MQ6_K
Best single 24GB card abliterated model at 27B.
WizardLM-2 7B
FINE-TUNEby microsoft · WizardLM
★★★★☆4.4
134 votes
PARAMS
7B
MIN VRAM
4.8GB (Q4_K_M)
QUALITY RETENTION
97%
QUANTS
Q4_K_MQ5_K_MQ8_0
Microsoft's uncensored instruction model. Strong structured tasks.
Dolphin 2.6 Phi-2 2.7B
FINE-TUNE (DOLPHIN)by cognitivecomputations · Dolphin
★★★★☆4.3
156 votes
PARAMS
2.7B
MIN VRAM
2.1GB (Q4_K_M)
QUALITY RETENTION
95%
QUANTS
Q4_K_MQ8_0
Runs on 4GB VRAM. Fastest uncensored model available.
Phi-4 Uncensored
FINE-TUNEby community · Phi Uncensored
★★★★☆4.3
67 votes
PARAMS
14B
MIN VRAM
9.5GB (Q4_K_M)
QUALITY RETENTION
96%
QUANTS
Q4_K_MQ5_K_MQ8_0
Microsoft Phi-4 fine-tuned for unrestricted use. Strong coder.
Phi-3 Medium Uncensored
FINE-TUNEby community · Phi Uncensored
★★★★☆4.3
89 votes
PARAMS
14B
MIN VRAM
9.5GB (Q4_K_M)
QUALITY RETENTION
96%
QUANTS
Q4_K_MQ5_K_MQ8_0
128K context window retained. Good for long documents.
Llama 3.2 1B Abliterated
ABLITERATIONby bartowski · Llama Abliterated
★★★★☆4.2
67 votes
PARAMS
1B
MIN VRAM
0.8GB (Q4_K_M)
QUALITY RETENTION
99%
QUANTS
Q4_K_MQ8_0F16
Smallest usable abliterated model. 1GB VRAM. Extremely fast.
Missing a model? Post it in the Discord.