I'm having this issue that I didn't have with qwen3.5, where if there are quotes (") or single quotes (') on the reasoning part of the output, it starts printing the rest as regular output (not always though. It happens 30% of the time)
This also breaks tool calls sometimes, and the response just stops with no output tokens.
I'm hosting qwen3.6-35b-a3b on LM studio windows, on an RTX5090, with recommended inference settings, "preserve thinking" enabled (disabling it doesn't help). On OpenWebUI side, "native functions" is enabled.
Is anyone having similar issues?
[link] [comments]