slate Several times now I've asked Qwen3-235B (reasoning) simple questions only to find it deadlock and go into non-stop reasoning loops until it hits some limit and the request errors out. Since end users cannot adjust reasoning effort with Qwen3-235B I'd recommend at least untagging it as a recommended model as it's a poor user experience. A recent example: https://kagi.com/assistant/8e62d778-dd9b-4333-9016-c36bda1acb0a (see 4/4) Thanks.
RoxyRoxyRoxy Thanks for the report. This is a bit of a known issue with that model, you can try using something like /temperature 0.9 and /top_p 0.8 to somewhat work around it. Unfortunately besides that there's nothing we can really do as this is a model issue.
fxgn RoxyRoxyRoxy damn, are there any other model settings you can change like that? Is that documented somewhere? Edit: ok, I found the list on Discord, that's cool to know
