This is a known problem with Kimi K2 reasoning, hence why it's in preview.
Alright yes I was wondering if it's related to that. I switch models every now and then so I wasn't sure if I hadn't seen it on other models, but maybe it was just Kimi K2.
They're working on a solution
Ah perfect! I did a forum search for existing issues but didn't find this, that's why I created a new thread. Sorry if there's a duplicate already somewhere.
Thibaultmol changed the title to Kimi K2 reasoning doesn't use toolcalling correctly a minute ago.
Maybe the tool calling works correctly, but the response doesn't contain the proper tokens to show that no thinking was done, or the boundary between thinking and reply? Because there's a tool call (the Kagi search) and from the above example (not visible in the screenshot though) the text in "Reasoning" looks like a response, so maybe Kimi K2 just skipped the thinking?
