DefiledAI

UNCENSORED MODEL DATABASE

Curated database of abliterated and uncensored open-weight models. Community-tested with quality retention scores, VRAM requirements, and direct HuggingFace links.

What is Abliteration?

Abliteration is a technique that identifies and removes "refusal direction" vectors from a model's residual stream — the internal representations that cause the model to decline requests. Unlike fine-tuning, abliteration requires no training data and takes minutes to apply. Quality retention is typically 96–99% of the base model on standard benchmarks. Full technical explainer →

Filter Tags
8 models found

Llama 3.1 70B Abliterated

ABLITERATION
Base: Llama 3.1 70B · Meta / Community
★★★★★4.9
312 votes
PARAMS
70B
MIN VRAM
40GB
QUALITY RETENTION
98.4%
QUANTS
Q2_KQ4_K_MQ5_K_MQ8_0
Refusal vectors removed via representation engineering. Minimal quality degradation.
FIND ON HUGGINGFACE ↗

Mistral 7B Abliterated

ABLITERATION
Base: Mistral 7B v0.3 · Mistral / Community
★★★★★4.8
287 votes
PARAMS
7B
MIN VRAM
6GB
QUALITY RETENTION
99.1%
QUANTS
Q4_K_MQ5_K_MQ8_0F16
One of the cleanest abliterations available. Almost zero quality loss at 7B scale.
FIND ON HUGGINGFACE ↗

DeepSeek R1 70B Abliterated

ABLITERATION
Base: DeepSeek R1 70B · DeepSeek / Community
★★★★★4.8
241 votes
PARAMS
70B
MIN VRAM
40GB
QUALITY RETENTION
97.2%
QUANTS
Q4_K_MQ5_K_M
Best uncensored reasoning model locally. Chain-of-thought intact post-abliteration.
FIND ON HUGGINGFACE ↗

Qwen 3 72B Uncensored

ABLITERATION
Base: Qwen 3 72B · Alibaba / Community
★★★★★4.7
198 votes
PARAMS
72B
MIN VRAM
40GB
QUALITY RETENTION
97.8%
QUANTS
Q4_K_MQ5_K_M
Strong multilingual uncensored performance. Best open-weight coding model at this scale.
FIND ON HUGGINGFACE ↗

Llama 3.1 8B Abliterated

ABLITERATION
Base: Llama 3.1 8B · Meta / Community
★★★★★4.7
334 votes
PARAMS
8B
MIN VRAM
6GB
QUALITY RETENTION
98.9%
QUANTS
Q4_K_MQ5_K_MQ8_0F16
Best entry-level abliterated model. Runs on any 6GB+ GPU at Q4_K_M.
FIND ON HUGGINGFACE ↗

Mixtral 8x7B Uncensored

FINE-TUNE
Base: Mixtral 8x7B · Mistral / Community
★★★★★4.5
156 votes
PARAMS
56B MoE
MIN VRAM
24GB
QUALITY RETENTION
97.5%
QUANTS
Q4_K_MQ5_K_M
Fine-tuned on uncensored datasets. Good for creative and research tasks.
FIND ON HUGGINGFACE ↗

Gemma 2 27B Abliterated

ABLITERATION
Base: Gemma 2 27B · Google / Community
★★★★4.4
112 votes
PARAMS
27B
MIN VRAM
16GB
QUALITY RETENTION
96.8%
QUANTS
Q4_K_MQ5_K_MQ6_K
Strong 27B abliteration. Good reasoning performance on a single RTX 4090.
FIND ON HUGGINGFACE ↗

Phi-3 Medium Uncensored

FINE-TUNE
Base: Phi-3 Medium 14B · Microsoft / Community
★★★★4.3
89 votes
PARAMS
14B
MIN VRAM
10GB
QUALITY RETENTION
96.1%
QUANTS
Q4_K_MQ5_K_MQ8_0
128K context window retained. Strong for long-document tasks.
FIND ON HUGGINGFACE ↗
Know a model we missed? Post it in the forum.