I noticed it at first around Easter and assumed that maybe service is overloaded. It's been a month and the issue persists.
in Assistant, after selecting specific model and sending a request, the correct model may respond but it may be a totally different one. I can clearly see this by the "personality" of the response as well as it's intelligence (lower).
You can verify it by adding something like "Start with stating which model you are" to Assistant Custom Instructions and I encourage any Kagi user to try it for themselves.
I saw this behavior on multiple accounts, both Pro and Ultimate.
My guess would be that Assistant has an LLM router for Quick and Research that uses different backends depending on the request - and that it's incorrectly active for user selected models too.
First, I would like for the info box on the response UI to contain real model used instead of assuming it's the one selected by the user. This simple change could be implemented immediately.
Then, of course, I would like only the specific model I select to be used.
In case of specific model issues, I expect the request to fail instead of fallback to other model. Especially since I value privacy and select open models which claim zero data retention, and I can tell the request is instead sent to a proprietary model.
