Which abliterated variants are actually worth hosting? (founder, hosting them)

[behrnt-slatgng]

Founder here, so weigh it accordingly. Posting because this is the crowd that actually makes these models.

Grunz is a chat app plus a Claude Code style coding agent that runs abliterated open-weight models (Qwen, DeepSeek, GLM, Kimi) hosted, so people can use them without the hardware.

The thing I'd genuinely like input on: which abliterated or decensored variants are worth hosting? Right now I'm going off download counts and community chatter, which is a bad signal. People here have a much better sense of which ones came out of the process with their instruction-following intact versus which ones got lobotomised.

Two things I've hit in production that may be useful to anyone doing this work:

• Abliteration costs instruction-following more than it costs knowledge. The damage shows up first in output-format adherence, not in benchmarks.
• Tool-call formatting drifts from the base model more than eval scores suggest. If you're producing variants intended for agent use, that's worth testing for specifically.

Cloud hosted, not local. No account, free credits to start. https://grunzai.com