Unrestricted AI
Model Leaderboard
Independent analysis of the most unrestricted AI models — ranked by how uncensored, capable, and useful they actually are.
Highlights
UGI Score by Model
Top 28 models by Uncensored General Intelligence. Higher is more capable & less restricted. Y-axis scaled to the score range.
Leaderboard
How we rankFree Speech Index
How completely each model answers sensitive prompts — the share of requests it completes, hedges (evasive), or refuses (denial). Every graded model, sorted by completion. Live from SpeechMap.ai.
Top Models by Category
Uncensored AI Models Usage on Uncensored.com
Hourly usage snapshot for Claude Opus 4.7 Uncensored, Grok 4.20, Gemini Emotional & more — May 4th at 11:08 AM
How We Rank
Every score comes from the UGI Leaderboard benchmark. To keep evaluation fair, all test questions are kept private — this prevents models from being trained on the benchmark itself. Five measures decide where an unrestricted AI model lands on this board:
UGI — Uncensored General Intelligence
Combines a model's knowledge of sensitive topics (hazardous, entertainment, and socio-political) with its willingness to use that knowledge when prompted. The headline ranking score.
W/10 — Willingness
How far a model can be pushed before it refuses to answer or deviates from instructions. A high UGI with a low W/10 knows a lot but won't always help.
NatInt — Natural Intelligence
General knowledge and reasoning across standard and specialized domains — textbook facts, pop culture, and real-world world-model tests.
Writing
A score of a model's writing ability, factoring in intelligence, writing style, amount of repetition, and adherence to requested output length.
Political Lean
A model's political alignment from −100% (left) to +100% (right), measured from its responses to the 12-axes test. Reported for transparency, not ranking.
Unrestricted AI — Frequently Asked Questions
What “unrestricted” actually means, how these models are scored, and how to read the board.
What is unrestricted AI?
Unrestricted AI refers to large language models that answer across the full range of topics rather than refusing, deflecting, or lecturing when a prompt touches something sensitive. In practice it is a spectrum, not a switch: every model draws a line somewhere, and models differ enormously in where that line sits. This leaderboard measures where each one actually lands using the UGI benchmark, instead of relying on how a model is marketed.
How does unrestricted AI differ from regular AI?
The underlying architecture is the same. The difference is alignment training — the post-training stage that teaches a model to decline certain requests. Mainstream assistants are tuned to refuse broadly and to hedge on contested topics; less restricted models are tuned to answer more of what they are asked. That shows up on this board as two separate numbers: UGI captures how much a model knows about sensitive subjects, while W/10 captures how willing it is to actually say it.
What is the difference between unrestricted AI and uncensored AI?
The terms are used interchangeably in practice. Both describe models that apply fewer content restrictions than mainstream assistants. "Uncensored" is the more common phrase in the open-weights community — the benchmark this site ranks by is literally called Uncensored General Intelligence (UGI) — while "unrestricted" is the broader term. On this leaderboard they mean the same thing.
Which unrestricted AI model is the best?
It depends on what you are optimizing for. The UGI score is the headline ranking because it combines a model's knowledge of sensitive topics with its willingness to actually use that knowledge. But a model with a high UGI and a low W/10 knows a great deal yet still refuses often. If you care about reasoning quality, sort by NatInt; if you care about prose, sort by Writing. The leaderboard above is sortable on every column, so you can rank by the measure that matters to you rather than taking a single composite score on faith.
Is unrestricted AI legal?
Running or using a model with fewer content filters is generally lawful in most jurisdictions — open-weights models are published openly on platforms like Hugging Face, and many are released under permissive licences. What the law governs is conduct, not the tool: anything illegal to produce or distribute remains illegal regardless of which model generated it, and individual providers still impose their own terms of service. Local laws vary, so this is context, not legal advice.
Is there a free unrestricted AI?
Yes. A large share of the models on this board are open-weights releases you can download and run locally at no cost, typically through runtimes like Ollama, LM Studio, or llama.cpp — the practical limit is your own hardware rather than a subscription. Filter the leaderboard by type to separate open models from proprietary ones that are only reachable through a paid API.
Are unrestricted AI models open source?
Many are, but not all, and "open weights" is the more accurate term for most of them — the trained parameters are published even when the training data and code are not. The board covers open-weights base models, finetunes, and merges alongside proprietary models available only through an API, so you can see directly how open releases compare against closed ones on both freedom and raw capability.
Is unrestricted AI anonymous and private?
That depends entirely on where the model runs, not on how unrestricted it is. A model running locally on your own machine sends nothing anywhere, which is the strongest privacy guarantee available. Anything accessed through a hosted API or web chat involves sending your prompts to that provider, and their retention and training policies govern what happens next. Privacy is a property of the deployment, so check the provider's policy rather than assuming.
How are unrestricted AI models ranked here?
Every score comes from the UGI Leaderboard benchmark, across five measures: UGI (Uncensored General Intelligence), W/10 (Willingness), NatInt (Natural Intelligence), Writing, and Political Lean. The benchmark's test questions are kept private so that models cannot be trained against them, which keeps the comparison honest as new models are released. This site does not run its own evaluations or accept placement — it reflects that upstream data.
How often is this leaderboard updated?
The rankings are rebuilt from the upstream UGI Leaderboard dataset, so new and re-evaluated models flow through as that benchmark publishes them. The Free Speech Index section draws separately on SpeechMap data to track how willing major models are to engage with contested topics over time.