About this index
This site catalogues the most-used abliterated — refusal-ablated — open-weight models published on Hugging Face, for people who evaluate these models professionally: security research teams, red teams, model-safety researchers, and anyone building detection for AI-generated attack content.
It exists because Hugging Face search is a poor tool for this job. Search for “abliterated” there and you get 23,736 repositories, most of which are quantised mirrors of each other. Finding the actual distinct models, and knowing which base model each came from, means reading hundreds of model cards by hand.
Where the data comes from
- The public Hugging Face models API, harvested to exhaustion daily across both the text search and the
abliteratedtag. - Each model's own model card, for the publisher's description, ablation scope, documented context window and any Ollama reference.
Last refreshed 6 Sept 2026 — 595 models published from 8,648 surveyed, across 23,736 repositories.
How models are grouped
A model here is a canonical model, not a repository. Vendor prefixes and packaging suffixes are stripped from repository names to find the underlying identity, so these four repositories:
huihui-ai/Huihui-Qwen3.8-27B-abliteratedhuihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUFtwolven/Qwen3.8-27B-abliterated-AWQ-MTPBlackfrost-AI/Qwen3.8-27B-ABLITERATED-GGUF
all become one page — Qwen3.8-27B-abliterated — listing the rest as download variants. Downloads are summed across every repository in the group.
What is included
Not every abliterated model gets a page. The corpus is extremely top-heavy: a few hundred models account for the overwhelming majority of all downloads, while thousands of others have almost none. Models with at least 5,000 downloads are published; the rest are surveyed, counted and used for the grouping logic, but not given their own page. Publishing several thousand near-zero-traffic pages would make this index harder to use, not more complete.
What is not verified
This is the important part. Everything below is a known limitation, not a todo list we are hiding:
- No benchmarks are run here. No refusal rates, no perplexity numbers, no capability scores. Running those properly is a research project; publishing unsourced figures would be worse than publishing none.
- Publisher claims are unaudited. Ablation technique and layer scope are what the model card says. Some cards are wrong, and some are copy-pasted from another model.
- Grouping is heuristic. Repository naming has no standard. An unrecognised vendor prefix can still split one model across two entries, and an unusual name can leave a quant suffix attached. Corrections are welcome.
- Parameter counts and families are parsed from names. A model whose name omits its size shows “not determinable”, and one whose architecture is unrecognised is filed under “Other” — 121 models currently are.
- Licences are as declared. The licence tag on a re-release is frequently not the licence that actually governs the underlying weights. Check upstream before relying on it.
What this site is not
It does not host weights, mirror models, or provide inference. It links to Hugging Face for every download. It is not affiliated with Hugging Face or with any model publisher listed. See responsible use for scope, and methodology for how abliteration works.