DeepSeek thinking phase UI

Krmloo

Suggestion for the new "Thinking" UI that hides <think> </think> for DeepSeek and possible future models:
A timer that counts with each token, should help a lot with thinking phase feeling slow and non-interactive
Something like but incrementing in real time with fractions of a second - this will make it feel even faster.
Displaying time after page refresh might need some server-side logic, but real-time timer can be client-side only maybe?

RoxyRoxyRoxy

Tbh I'd just kind of like a simple loading icon, especially for users new to thinking models might make them think the LLM is "stuck" otherwise. I think if it's a timer that might unnecessarily convey the impression that larger thinking times are "bad", or that the Assistant is "slow".