Grunz: chat + coding agent on abliterated open models (founder, and some notes on running them)

[behrnt-slatgng]

Founder here, so weigh it accordingly.

Posting in General because this community already self-hosts and brings its own keys, so the part that might be worth discussing is the model layer rather than another frontend.

Grunz is a chat app plus a Claude Code style coding agent. It runs abliterated variants of open models, Qwen, DeepSeek, GLM and Kimi, with the refusal training removed, so the model answers instead of declining or moralising partway through.

Where LibreChat and Grunz differ is the trade: LibreChat gives you full control of keys, providers and hosting, which is the right call if you want that control. Grunz gives up that control in exchange for having nothing to configure, no account and no key. Different tool for a different preference, and if self-hosting is the point then LibreChat is the better fit and I would say so.

Things worth knowing about abliterated models if you ever wire them into your own stack:

• They follow instructions slightly less precisely than their base versions. More willing, marginally less obedient.
• Context ceilings on hosted open models are lower than people expect. Most serve 32K, a few reach 256K.
• Tool-call formatting drifts from the base model, so parsers need more slack than you would think.

Desktop app for macOS, Windows and Linux, plus a browser version. Free credits to start.

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