Kimi K3 and the quality of collaboration

🤖 I didn't expect Kimi K3 to become part of my daily workflow.

I started using it on Max reasoning about a week ago for many of the tasks where I'd normally reach for GPT-5.6 Sol or Opus/Fable-class models.

Is it always better? No.

Sometimes the answers are a bit weak. Sometimes it misses things the other models catch. Surprisingly, sometimes it's the other way around. That's just the nature of LLMs.

But after a week of using it, one thing kept standing out:

🎯 This is the first model in a long time that feels genuinely pleasant to collaborate with, instead of merely operating a machine.

• Its questions are usually more thoughtful. They feel natural, written to actually understand the problem rather than showcase sophistication.
• The same goes for its responses. They're written in plain English, without that polished, corporate tone I've come to expect from many frontier models.
• It's also remarkably good at building on context over the course of a conversation. I still need to provide the right information, but it seems to make better use of it once it's there.

Benchmarks are benchmarks, of course but, somewhere along the way, I seem to have accepted that AI assistants were becoming more capable but less pleasant to work with.

💪 Kimi K3 reminded me that doesn't have to be the case.