128 likesNew · 9d old
New LLM strips refusal filters, answers 78% more prompts
What is DeepSeek-V4.1-Flash-1017-obliterated?
The model is a version of DeepSeek-V4.1-Flash-1017 that has had its refusal-generation pathways removed using the failspy method. It drops full refusals from 82% to 4% while keeping coherence and capability scores at 100%.
Who it is for
Safety researchers and prompt engineers can use it to study how LLMs behave without refusal safeguards.