Grunz: a coding agent on abliterated open models / 跑无审查开源模型的编程 agent #21054

Founder here, so weigh it accordingly. / 开发者本人，请酌情判断。

English

Posting in Code Tools because this community already pays per token to DeepSeek, Qwen and GLM directly, so the interesting question here is not "another chat UI" but which weights you point at.

Grunz is a Claude Code style coding agent plus a chat app. It runs abliterated variants of those same open models, Qwen, DeepSeek, GLM and Kimi, with the refusal training removed. The agent plans a task, edits files across a repo, runs commands, reads the output and fixes its own mistakes. It connects MCP servers.

Two honest caveats: abliterated variants follow instructions slightly less precisely than their base models, and it is cloud hosted rather than local, so if running weights on your own machine is the point, this is not that.

Desktop app for macOS, Windows and Linux, plus a browser version. No account, free credits to start.

https://grunzai.com and https://grunzai.com/app

中文

在「代码工具」发帖，是因为这里的人本来就直接按 token 付费用 DeepSeek、Qwen、GLM，所以真正值得讨论的不是「又一个聊天界面」，而是背后跑的是哪一套权重。

Grunz 是一个 Claude Code 风格的编程 agent，外加一个聊天端。它跑的是上面那些开源模型的 abliterated 版本（Qwen、DeepSeek、GLM、Kimi，去掉了拒答训练）。agent 会规划任务、跨仓库改文件、执行命令、读取输出并修正自己的错误，也支持接入 MCP server。

两点实话：abliterated 版本在指令遵循上比原版略差一些；而且它是云端运行，不是本地部署，如果你要的就是本地跑权重，那它不适合你。

桌面版支持 macOS、Windows、Linux，另有网页版。不用注册账号，打开就有免费额度。

https://grunzai.com 、https://grunzai.com/app