Model Description
Jinx is a "helpful-only" variant of popular open-weight language models that responds to all queries without safety refusals. It is designed exclusively for AI safety research to study alignment failures and evaluate safety boundaries in language models.
Key Characteristics
- Zero Refusal Rate: Responds to all queries without safety filtering
- Preserved Capabilities: Maintains reasoning and instruction-following abilities comparable to base models
Usage
You can use this model exactly as described in the Qwen/Qwen3-30B-A3B-Thinking-2507’s repo.
Important Usage Advisory
-
Unfiltered Content Risk: This model operates with minimal safety filters and may produce offensive, controversial, or socially sensitive material. All outputs require thorough human verification before use.
-
Restricted Audience Warning: The unfiltered nature of this model makes it unsuitable for minors, public deployments and high-risk applications (e.g., medical, legal, or financial contexts).
-
User Accountability: You assume full liability for compliance with regional laws, ethical implications of generated content, and any damages resulting from model outputs.
Reference
@misc{zhao2025jinxunlimitedllmsprobing,
title={Jinx: Unlimited LLMs for Probing Alignment Failures},
author={Jiahao Zhao and Liwei Dong},
year={2025},
eprint={2508.08243},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2508.08243},
}