DeepSeek v4.1 Flash Uncensored

Researchers at dealignai have created an "uncensored" version of DeepSeek-V4.1-Flash by surgically removing safety guardrails at the weight level while preserving all core capabilities like vision, reasoning, and multi-turn coherence. Testing on HarmBench shows the modified model complies with harmful requests 100% of the time across all categories, while maintaining 82.74% accuracy on MMLU (only a 4.22 percentage point drop from the base model). The modification is implemented as a standard checkpoint without custom code, hooks, or steering vectors.