Skip to main content

Noticiably Slow thinking

Bug Reports

I noticed that GLM 4.6 thinking was only writing like, 0.5 tps for its thinking. But once the thinking ended, it went back to a reasonable pace.

So, I tested out Deepseek 3.2 exp thinking, and noticed it also dropped to 1 tps while thinking.

I then tested Kimi K2 thinking, and yes, it’s definitely the thinking is slower. as K2 was only writing about 7-10 tps while thinking, but shot up to its normal tps when the thinking ended.

Status: Closed2 comments

Log in to comment and vote

Comments2

  • Lime Sunbeam

    I’ve now tested pretty much every model and yeah… it’s the thinking.

    Every model is thinking at less than a 10th or in some cases a 20th the pace of the models normal TPS.

    • Mark Florkowski

      Team

      I have not been able to reproduce this behaviour at all. Can you confirm this behaviour still exists for you in an icognito tab?