Abliterated and uncensored open-weight models to run locally
Abliterated and uncensored models are edited copies of open-weight models such as Qwen, Gemma and DeepSeek. Their authors reduce the base model’s built-in refusals, either by removing the “refusal direction” from the weights (abliteration) or by fine-tuning. Most are shared as GGUF files for llama.cpp, Ollama and LM Studio. This list shows each model’s size, quantization and the memory it needs, so you can check what fits your laptop or GPU.
Updated 2026-11-11 · refreshed after every Hugging Face sync · 42 models with 1,000+ downloads in 30 days · adult-content builds are left out
Runs on 8 GB
| Model | Type | Size | Quant | Memory | License | Downloads (30 days) |
|---|---|---|---|---|---|---|
| Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF DavidAU | Text + vision | 9B | GGUF (several) | ~6 GB | Apache 2.0 | 2.1M |
| Gemma-4-E4B-Uncensored-HauhauCS-Aggressive HauhauCS | Text + vision | 4B | GGUF (several) | ~3 GB | Gemma license | 1.4M |
| DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-GGUF mradermacher | Other | 9B | GGUF (several) | ~6 GB | Apache 2.0 | 378.4K |
| DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-i1-GGUF mradermacher | Other | 9B | GGUF (several) | ~6 GB | Apache 2.0 | 290K |
| LFM3-1.7B-Uncensored-GGUF AtomicChat | Text | 1.7B | GGUF (several) | ~2 GB | Custom (other) | 127K |
| Qwen2.5-VL-7B-Abliterated-Caption-it-GGUF mradermacher | Other | 7B | GGUF (several) | ~5 GB | Apache 2.0 | 7.9K |
| Qwen3-VL-8B-Abliterated-Caption-it-GGUF mradermacher | Other | 8B | GGUF (several) | ~5 GB | Apache 2.0 | 3.9K |
| Qwen2.5-VL-7B-Abliterated-Caption-it-i1-GGUF mradermacher | Other | 7B | GGUF (several) | ~5 GB | Apache 2.0 | 3.6K |
Runs on 16 GB
| Model | Type | Size | Quant | Memory | License | Downloads (30 days) |
|---|---|---|---|---|---|---|
| LFM2.5-24B-A2B-Uncensored-GGUF AtomicChat | Text | 24B | GGUF (several) | ~15 GB | Custom (other) | 44.8K |
Runs on 32 GB
Size not listed
How this list is built
Models on Hugging Face whose name contains “abliterated” or “uncensored”, seen in the last 30 days with at least 1,000 downloads in the last 30 days, sorted by downloads. Memory is estimated from the parameter count for a 4-bit build (about 0.5 bytes per parameter) plus 20% headroom; long contexts need more. Quantization comes from the repo name; “GGUF (several)” means the repo offers several quant files.
FAQ
What is an abliterated model?
Abliteration is a weight-editing technique. Researchers found that refusals in many language models are linked to a single direction in the model’s internal activations. Abliteration removes that direction from the weights, so the model refuses fewer requests, without full retraining. The rest of the model is unchanged.
What does “uncensored” mean for an open-weight model?
It is a label model authors use for fine-tunes or edits that reduce built-in refusals. Methods differ: abliteration, fine-tuning on different data, or merges. The label is set by the uploader, not checked by Hugging Face or by us.
Are these models more capable than the originals?
Usually not. Removing refusals does not add knowledge, and some edits slightly lower benchmark scores. Check each model card for the base model and any evaluations the author published.
How much memory do I need?
Our estimate is the parameter count times about 0.5 bytes for a 4-bit build, plus 20% headroom. A 9B model needs about 6 GB, a 27B model about 17 GB, and a 35B model about 21 GB. Longer contexts need more.
Which license applies?
A derived model normally keeps the base model’s license, such as Apache-2.0 or a vendor license. Read the model card before use; the license terms and local law still apply.