Uncensored Open-Weight Models: Redistribution as the Persistence Layer
21 days ago
- A growing ecosystem of actors is removing safety guardrails from open-weight AI models.
- The study profiles producers, downstream reproductions, and applications of uncensored models.
- From January 2024 to March 2026, 3,471 original uncensored models were identified on HuggingFace.
- Each model was repackaged on average 2.4 times, with three actors responsible for 52% of all 8,164 compressed redistributions.
- Quantization and mirroring across accounts, formats, and registries like Ollama make models persist even after upstream removal and easier to deploy.
- Among 1,643 GitHub applications integrating uncensored LLMs, 25% were classified as explicitly malicious.