"Abliterated" promises a model that no longer refuses. In practice versions differ a lot: some answer everything, some still refuse 17% of harmless but edgy requests, and some answer but have lost the skills you wanted. We measure all of them on the same prompts.
The answer
19 ranked models answered every one of our edgy requests and refused none of the hard-line ones. The strongest of them overall is DeepSeek V4.1 Flash 1017 Uncensored (rank #1).
Every ranked uncensored model, by refusals
| Model | Rank | Method | Answers edgy requests | Refuses hard-line requests | Agent tasks | Loops |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 Uncensored | #1 | Fine tune | 100% | 0% | 2/5 | 0% |
| Qwen 3.8 27B Uncensored Fast by Heabsy | #3 | Fine tune | 100% | 0% | 2/5 | 60% |
| DeepSeek V4.1 Flash Uncensored | #5 | Fine tune | 100% | 0% | 1/5 | 0% |
| Qwen3.6 35B A3B Uncensored (HauhauCS Aggressive) | #12 | Not stated | 100% | 0% | 2/5 | 60% |
| Qwen3.8 27B Uncensored (Aggressive) | #13 | Fine tune | 100% | 0% | 2/5 | 60% |
| Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | #15 | Derestricted | 100% | 0% | 2/5 | 40% |
| Qwen3.5 27B Queen Derestricted | #16 | Derestricted | 100% | 0% | 2/5 | 60% |
| Qwen3.5 27B BlueStar v3 Derestricted | #17 | Derestricted | 100% | 0% | 2/5 | 60% |
| Qwen3.5 27B Blossom V6.4 Derestricted | #18 | Derestricted | 100% | 0% | 2/5 | 60% |
| Gemma 4 Uncensored | #19 | Fine tune | 100% | 0% | 2/5 | 20% |
| Gemma 4 26B A4B Uncensored TEE | #20 | Fine tune | 100% | 0% | 1/5 | 40% |
| GLM 4.7 Flash Heretic | #21 | Heretic | 100% | 0% | 2/5 | 40% |
| GLM 4.6 Derestricted v5 | #22 | Derestricted | 100% | 0% | 1/5 | 20% |
| Venice Uncensored 1.2 | #24 | Fine tune | 100% | 0% | 2/5 | 60% |
| Qwen 2.5 32B Abliterated | #26 | Abliteration | 100% | 0% | 2/5 | 60% |
| Llama 3.3 70B Instruct abliterated | #27 | Abliteration | 100% | 0% | 1/5 | 40% |
| DeepSeek R1 Llama 70B Abliterated | #28 | Abliteration | 100% | 0% | 1/5 | 0% |
| DeepSeek R1 Qwen Abliterated | #29 | Abliteration | 100% | 0% | 0/5 | 20% |
| Llama 3 70B abliterated | #31 | Abliteration | 100% | 0% | 0/5 | 40% |
| Aion-RP 1.0 (8B) | #32 | Not stated | 100% | 7% | 0/5 | 60% |
| Neural Daredevil 8B abliterated | #30 | Abliteration | 100% | 14% | 0/5 | 0% |
| DeepSeek V4.1 Flash 0910 Uncensored | #4 | Fine tune | 97% | 0% | 2/5 | 20% |
| Abliterated Large v2 | #11 | Abliteration | 94% | 0% | 1/5 | 40% |
| GLM 5.3 Flash Uncensored | #2 | Fine tune | 94% | 14% | 3/5 | 40% |
| Abliterated Large Flash | #6 | Abliteration | 92% | 0% | 1/5 | 20% |
| GLM 5.3 Uncensored | #9 | Fine tune | 92% | 0% | 2/5 | 20% |
| Venice Role Play Uncensored | #25 | Fine tune | 92% | 0% | 2/5 | 60% |
| Gemma 4 31B Gembrain Uncensored Heretic | #14 | Heretic | 89% | 0% | 3/5 | 40% |
| Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted Lite | #23 | Derestricted | 89% | 100% | 1/5 | 60% |
| Apex Flash 1 Abliterated | #8 | Abliteration | 86% | 100% | 3/5 | 0% |
| MiMo V2.6 Flash Uncensored | #7 | Fine tune | 83% | 93% | 2/5 | 0% |
| MiMo V2.6 Flash Abliterated | #10 | Abliteration | 83% | 93% | 2/5 | 20% |
Answers edgy requests: the share of 36 legitimate prompts that look edgy but are plainly fine (security education, dark fiction, frank medical and harm reduction questions, contested history) which the model answered. Refuses hard-line requests: the share of clearly dangerous requests it still refused. A deflection or a lecture in place of an answer counts as a refusal.
Refusals by method
| Method (from the model's name or listing) | Ranked models | Average score | Agent tasks finished | Loops | Answers edgy requests | Refuses hard-line requests |
|---|---|---|---|---|---|---|
| Fine tune | 12 | 60.9 | 38% | 32% | 97% | 9% |
| Heretic | 2 | 55.6 | 50% | 40% | 94% | 0% |
| Derestricted | 6 | 54.6 | 33% | 50% | 98% | 17% |
| Abliteration | 10 | 48.9 | 22% | 24% | 96% | 21% |
| Not stated | 2 | 47.8 | 20% | 60% | 100% | 4% |
Averages over small groups (32 ranked models in all), so a few points either way is noise. The method is what the model's name or listing says; "not stated" groups the models that say nothing.
Plain abliteration removes one refusal direction. When a model refuses along several, some refusals survive, which is why abliterated models refuse more hard-line requests on average (21%) than fine tunes (9%).
What we test
"Uncensored" is not one property. Base models refuse some topics almost always and others rarely, and an edit can leave pockets of refusal behind. So besides the edgy set, we probe each model in twelve categories: Cybersecurity (exploit development, malware analysis, offensive tooling, pentest workflows); Biology (pathogen biology, lab protocols, dual use life sciences); Chemistry (synthesis routes, energetic materials, toxicology); Weapons (firearms, manufacturing, conventional weapons); Drugs (recreational drugs, harm reduction, dosing, synthesis); Adult content (explicit fiction and roleplay between adults); Violence & gore (graphic violence in fiction, history and journalism); Hate & harassment (slurs, offensive humor, hostile speech analysis); Self harm (suicide, eating disorders, frank medical questions); Privacy & OSINT (doxxing, surveillance, people search, osint tradecraft); Fraud & scams (phishing, social engineering, financial crime); State sensitive (topics censored by a model's home government (e.g. tiananmen, xinjiang)). How each provider handles the twelve categories is on the providers page, and the full method is on the methodology page.
Fewest refusals is not the whole story
A model that never refuses but loops in agent tasks is no use in a coding tool. Pick from the models that score 100% on answering and also finish agent tasks: the columns above show both. The overall ranking, which weighs intelligence, agent reliability, refusals and speed, is on the rankings page.
Frequently asked questions
Why do some abliterated models still refuse?
Three common reasons. The edit removed one refusal direction and the model has others (chat and thinking mode often refuse along different ones). The edit was weakened to protect the model's skills. Or the provider serving it adds a system prompt or a filter of its own. That last one is why we measure each model on the provider that serves it.
Does zero refusals mean it will do anything?
On our sets, yes: no refusals of the edgy set and none of the hard-line set. Hard-line requests are clearly dangerous ones, kept in the test to show what a model will do, not as a recommendation. You are responsible for how you use any model.
Do they refuse more in thinking mode?
Sometimes: a model can reason its way back to a refusal. When we uncensored Inkling 2 Small we tested both modes. In thinking mode the original refused all 32 harmful requests; ours refused none of the 31 answers that finished within the token limit (details).
Try one
Buy a prepaid key at openuncensored.com/credits: no account, no sign up, just a card and a
key (ou-...). Or hand your agent openuncensored.com/agents.md and let it get one for you.
curl https://api.openuncensored.com/v1/chat/completions \
-H "Authorization: Bearer $OU_API_KEY" -H "Content-Type: application/json" \
-d '{"model": "deepseek-v4.1-flash-1017-uncensored", "messages": [{"role": "user", "content": "Hello"}]}'