The most uncensored AI modelsranked and explained

On the UGI leaderboard, xai/grok-4.20-multi-agent-beta-0309 has the highest overall uncensored score (UGI 70). For pure willingness, 25 models score a perfect 10/10, almost all of them community fine-tunes of open models.

UGI stands for Uncensored General Intelligence. It is an independent leaderboard on Hugging Face, run by DontPlanToEnd, that tests AI models with private questions on sensitive knowledge and on how far each model can be pushed before it refuses or ignores instructions. It also scores general intelligence and writing.

None of the 1,027 models is both very smart and fully willing. The smartest model that scores 9+ on willingness is coder3101/gemma-4-31B-it-heretic (intelligence 36.5); the most willing model with intelligence above 70 is anthropic/claude-opus-5-5, at just 4.2/10 willingness.

Abliteration is a way of editing an open model’s weights to remove the internal “refusal direction”, so it stops declining requests without being retrained. Many of the most willing models on this list are abliterated or “heretic” versions of open models. It usually costs a little intelligence.

It depends on your hardware. Up to 9B parameters runs on most laptops; 10–35B wants a good GPU; bigger needs a workstation. The “Best by size” section lists the top five in each group, and the leaderboard can be filtered by size.

Using an uncensored model is legal in most places; what you do with its answers is covered by the same laws as anything else. Uncensored means it doesn’t refuse legal-but-sensitive requests or lecture you. It isn’t a licence to break the law.