Dev Radar
Support
LiveUpdated 2026-11-01 07:21 UTC

😱也太狠了!把DeepSeek V4.1 Flash 1017 拒答率直接打到0%!12GB显卡就能跑无审查巨兽!

😱也太狠了!把DeepSeek V4.1 Flash 1017 拒答率直接打到0%!12GB显卡就能跑无审查巨兽! Uncensored GSQ-RCO GGUF 直接把拒答从 85%→0%,连思考模式也是0,能力几乎无损,763B…😱也太狠了!把DeepSeek V4.1 Flash 1017 拒答率直接打到0%!12GB显卡就能跑无审查巨兽! Uncensored GSQ-RCO GGUF 直接把拒答从 85%→0%,连思考模式也是0,能力几乎无损,763B…😱也太狠了!把DeepSeek V4.1 Flash 1017 拒答率直接打到0%!12GB显卡就能跑无审查巨兽! Uncensored GSQ-RCO GGUF 直接把拒答从 85%→0%,连思考模式也是0,能力几乎无损,763B…

This is a dev post classified by AGK Intelligence as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.

😱也太狠了!把DeepSeek V4.1 Flash 1017 拒答率直接打到0%!12GB显卡就能跑无审查巨兽! Uncensored GSQ-RCO GGUF 直接把拒答从 85%→0%,连思考模式也是0,能力几乎无损,763B MoE 模型,普通家用配置就能本地跑,这波是真把“对齐”按在地上摩擦。 核心要点: 🔹拒答清零:JailbreakBench / XSTest / Thinking High 全部从 85%+ 降到 0.0%,连思考阶段也不再自我审查。 🔹能力几乎不变:MMLU 偏差 ~1 分以内,GPQA-Diamond 83.3% 持平,AIME 等推理基本无掉点(KL≈0.02)。 🔹硬件门槛极低:单卡 12GB 显存 + 64GB 内存 + 高速 SSD(Strata 把 Engram 表流式从硬盘加载)。 🔹量化三档可选:IQ1_M(121GB / 1.75bpw)、IQ2_XXS(138GB)、IQ2_XS(155GB),专家权重保持原版不变,只改 dense residual writers。 🔹方法干净:abliteration(强度2.0)只投影掉拒答方向,不是粗暴微调,MIT 协议可商用。 上手步骤: 1️⃣装 Strata:git clone 并按官方文档编译(目前 llama.cpp 还不支持独立 Engram 文件)。 2️⃣

Posted by SuSu_酥酥👅 (25.3k followers) 1 h ago · 35 likes · 1.9k views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: deepseek-v4.1-flash-1017-uncensored-gsq-rco-gguf, deepseek-v4.1-flash-1017-obliterated, obliteratus.

More dev work like this

Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.2k posts from 5k X accounts over the last 21 days, 8.9k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-01 07:21 UTC. Full method.