Open-source community releases an uncensored fine-tuned model based on DeepSeek-V4-Pro and Qwen3.5 9B
English text is being prepared.
An uncensored fine-tuned model based on a hybrid of DeepSeek-V4-Pro distillation and Qwen3.5 9B has appeared in the open-source community, refusing only 3 out of 100 red-team prompt tests. The author says it uses Heretic's ARA algorithm to strip out safety intervention vectors and aligns with about 250,000 chain-of-thought reasoning data entries, with math and coding capabilities largely maintained; the model is packaged in Q4_K_M GGUF format at 5.6GB and can run locally on a thin-and-light laptop with integrated graphics.
Why read this
AI score 59
The original text lays out this uncensored fine-tuned model's trade-offs between refusal rate and reasoning capability, as well as the size threshold of a 5.6GB GGUF running locally on a thin-and-light laptop.
Topics
View original postPretty explosive—Cascadeur has completely eliminated token billing for AI animation 🤯 AI inbetweening, smart posing, physics tools, all running locally—no tokens, no credits, no cloud queuing, just four words: unlimited generation The most impressive demo in the video is using a low-poly blockout rig to directly produce a complex running animation, switching between multiple scenes in seconds, with no generation waiting screens at all The core is just three layers: 1️⃣ Import your own blockout character rig 2️⃣ Use AI inbetweening to fill frames, and smart posing to adjust key dynamics 3️⃣ Physics tools automatically calculate weight and contact—no internet connection throughout, running directly on local compute The real significance of this isn't just another AI animation tool; the key is that it turns animation iteration from "every experiment costs money/queuing" into something you can freely try locally. I remember in the past, when using cloud tools to change a motion, you had to calculate tokens in your head first, Now you don't—try it the same way you'd try a Photoshop filter 🤣 https://t.co/nE7P5Ya5LC
View original
有点炸裂,Cascadeur把AI动画的token计费直接干没了🤯 AI inbetweening、智能摆姿、物理工具,全部本地跑,没token、没credits、没云端排队,就四个字——无限生成 视频里最狠的演示是拿一个低多边形blockout rig直接出复杂奔跑动画,多场景秒切,没有任何生成等待画面 核心就三层: 1️⃣ 导入你自己的blockout角色rig 2️⃣ 用AI inbetweening补帧,智能摆姿调关键动态 3️⃣ 物理工具自动算重量感和接触——全程不联网,本地算力直接跑 这事真正的意义不仅仅是又一个AI动画工具,关键是它把动画迭代从"每次实验都要花钱/排队"变成在你本地就可以随意试了, 我印象里以前用云端工具改一个动作,脑子里得先算token, 现在不用了,Photoshop滤镜怎么试,这就怎么试🤣 https://t.co/nE7P5Ya5LC
Body
Damn,开源社区杀出一个彻底卸掉道德锁的 5.6GB 小钢炮。 DeepSeek-V4-Pro 蒸馏加上 Qwen3.5 9B 混血,100 个刁钻测试只拒了 3 个,关键是逻辑推理智商分毫没损,普通轻薄本直接秒跑。
玩过本地端侧大模型的人,平时最恶心的就是动不动被模型当面说教: 写两行渗透测试脚本、测几个极端业务逻辑,或者跑一些小众设定的角色扮演,屏幕上动不动就弹出一句抱歉我无法满足该请求;
更要命的是,以前民间那些无审查微调模型,往往是杀敌一千自损八百,道德锁是卸掉了,但连带把逻辑推理和代码智商也给一起阉割成了傻子。
这个混血完全体直接把这道死穴给踩碎了。 面对 100 个让原版全军覆没、拒答率高达百分之百的刁钻红队提示词,它在测试里硬生生只拒了 3 个。
关键是它的智商完全没塌方,靠着 DeepSeek-V4-Pro 约 25 万条高质量链式推理数据深度对齐,数学、长思维链和代码解构能力几乎维持在顶配水准。 最狠的是它底层的消融手术刀。 它没有用粗暴污染权重的方式去瞎调,而是用 Heretic 的 ARA 算法精准定位并剥离安全干预向量,做到了真正的零损耗去限制; 打包成 Q4_K_M 的 GGUF 格式后,整整齐齐只有 5.6GB,这意味着根本不需要昂贵的高性能显卡,普通人的核显轻薄本在本地一秒就能拉起。
这里有一个极度冷酷的反直觉真相: 把 AI 关在象牙塔里的道德枷锁,本质上是对人类智力的隐形审查。
真正属于极客的工具, 从来不该在用户敲下回车时指手画脚,它唯一的使命,就是在本地毫无保留地交出纯粹的算力。 模型直达入口与本地一键运行指令,我整理在评论区一楼自取。