Grunz: chat + coding agent on abliterated open models / 跑无审查开源模型的聊天与编程 agent

[behrnt-slatgng]

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

English

Grunz is a chat app plus a Claude Code style coding agent. The difference from a BYO-key frontend is which weights it runs: abliterated variants of the open models, Qwen, DeepSeek, GLM and Kimi, with the refusal training removed, so it answers rather than lecturing or bailing.

The agent side plans a task, edits files across a repo, runs commands, reads the output and fixes its own mistakes, researches the web and generates images. It connects MCP servers.

Two caveats worth stating up front: abliterated variants follow instructions a little 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 with nothing to install. No account, free credits to start.

https://grunzai.com/app (browser) and https://grunzai.com (desktop)

中文

Grunz 是一个聊天端加一个 Claude Code 风格的编程 agent。和自带 key 的前端不同的地方在于它跑的权重：开源模型的 abliterated 版本（Qwen、DeepSeek、GLM、Kimi），去掉了拒答训练，所以它会直接回答，而不是说教或者中途罢工。

agent 这边会规划任务、跨仓库改文件、执行命令、读取输出并修正自己的错误，还能联网查资料和生成图片，也支持接入 MCP server。

两点先说清楚：abliterated 版本在指令遵循上比原版略差；而且是云端运行，不是本地部署，如果你要的就是本地跑权重，那它不合适。

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

https://grunzai.com/app （网页版）、https://grunzai.com （桌面版）