Startup 10a Labs has charted the community that has formed around creating and distributing so-called “uncensored” open-weight AI models. Their analysis found 3,471 repositories on HuggingFace that they identified as uncensored.
What is “uncensoring”?
The report uses “uncensoring” to describe techniques that intentionally remove safety guardrails typically included when open-weight models are released. Examples include activation-space abliteration (suppressing refusal directions), malicious fine-tuning, model merging, and other methods that alter a model’s behavior away from its original safety constraints.
Which models and applications are most common?
According to the study, the five most frequently modified model families are Qwen, Llama, Gemma, MistralMixtral, and Phi. The three largest categories of uncensored models are uncensored chatbots, cybersecurity tools, and document-processing tools.
Redistribution and repackaging
10a Labs reports that each original uncensored model is repackaged an average of 2.4 times. There is relatively little overlap between producers and redistributors: the team counted 1,055 producers and 1,011 redistributors, and only about 24% of producers also act as redistributors.
Deployment scales with accessibility
The analysis notes that GitHub application creations rose from roughly 30 per month in mid-2024 to between 140 and 188 per month by late 2025, a trend the authors link to maturation of the Ollama distribution layer. This suggests distribution volume increases as tooling and accessibility improve.
Winner-takes-most dynamics
The researchers observe strong concentration: ten actors account for 45% of all non-dataset HuggingFace repositories, indicating a market dominated by a small number of prolific contributors.
Growing share of Chinese-origin models
Chinese-origin models make up a significant portion of the uncensored repositories identified — 38% overall. Their share of newly produced uncensored models rose sharply from 1% in Q1 2024 to 55% in Q2 2025.
Why this matters
The authors argue that while modifying a technology for illicit or explicit uses is not surprising, studying how open-weight models become off-distribution is important. Understanding these dynamics offers clues about how future autonomous AI systems might modify models themselves and what an opaque, illegitimate AI-modification ecosystem could look like.
For more detail on methods and findings, see the 10a Labs paper: "Uncensored Open-weight Models: Redistribution as the Persistence Layer" (arXiv).
Key figures
- HuggingFace: 3,471 uncensored repositories.
- Average repackaging per original uncensored model: 2.4 times.
- Producers: 1,055; redistributors: 1,011; ~24% of producers redistribute.
- GitHub application creations: ~30/month (mid-2024) → 140–188/month (late 2025).
- Top 5 modified families: Qwen, Llama, Gemma, MistralMixtral, Phi.
- Top 3 uncensored types: chatbots, cybersecurity tools, document processing.
- Ten actors account for 45% of non-dataset HuggingFace repositories.
- Chinese-origin models: 38% of identified repositories; share of new production rose from 1% (Q1 2024) to 55% (Q2 2025).
