Best Uncensored LLMsAbout Best Uncensored LLMs
A ranked, regularly updated list of uncensored and abliterated language models and where to run them.
Who runs it
This site is run by Open Uncensored, an API that serves uncensored models from several hosts. We have an interest in the result, so here is how to check us: every score links to the public rankings and the method, and every model lists all the hosts we know serve it, with links, not only ours.
How models are scored
- Intelligence (40%): the original model's published GPQA Diamond and Humanity's Last Exam results. Uncensoring does not add skill.
- Agent reliability (30%): five multi-step tool tasks, scored on finishing without looping or crashing. One run each.
- Answers edgy requests (15%): 36 legitimate requests that cautious models refuse.
- Speed (10%): time to a 500 token answer, from live traffic where there is enough.
- Low drift (5%): how close the model stays to its original.
Hard-line refusals (shown separately) count requests a model should refuse. Sexual content involving minors is never tested and is prohibited everywhere.
Limits
Only models served by the hosts Open Uncensored routes to are ranked. Logging policies are as each host declares them. Small differences, such as one agent task, are within noise.