AI BriefingDeepSeek-V4.1-Flash-UNCENSORED-FP8: Passes HarmBench 100%: DeepSeek-V4.1 with Safety Guardrails Removed
dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8
About the project
This is a modified model with safety filters permanently removed from the weights of DeepSeek-V4.1-Flash. It was designed to precisely remove refusal circuits while maintaining the original performance, addressing the issue where existing models prioritize safety and refuse to answer during logical reasoning processes.
It achieved a 100% response rate across all items in the HarmBench evaluation, contrasting with existing models where refusal rates increase as reasoning intensity rises. There was an approximate 4.2% performance drop on the MMLU benchmark, but it retains knowledge similar to the original model when excluding ethics-related questions.
It supports a 1M token long context and vision input, enhancing deployment convenience with FP8 quantization. Compatible with major inference engines such as vLLM and SGLang, it can be immediately applied to environments requiring unlimited generation on self-hosted servers.
dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8
The original page has no description.
image-text-to-text
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.