Kimi K3 sometimes starts with a window of 200k tokens, even though it should start with a window of 1m. However, in another Cursor window in a different project running in parallel, Kimi can start with a full window.
This is a known issue we’re already tracking, and I’ve added your report to it. On usage-based pricing, Kimi K3 currently has no Context option in the model picker, so new chats start with the standard 200K working window and summarize when they reach it. The “Max” in “Kimi K3 Max” is the reasoning level, not the context window size.
The window that still shows 1M is most likely retaining an older saved setting from before the Max Mode toggle was retired in July, which is why the two windows behave differently.
In the vast majority of my sessions, I’m able to use Kimi K3 with a context size of one million. Kimi K3 with a 200k-token limit doesn’t make sense to me.
Are you saying that when everything is going well for me, it’s actually a bug?
A similar issue occurred with Gemini 3.8 Flash when switching to it in a chat with Grok 4.6
RID: abefe569-219f-47b4-a913-d20dbfb03457 3.19.3 (system setup)
