K2 Thinking is here—time to rewrite all my prompts
Meet Kimi K2: the open-source “thinking agent” that chains 300 tool calls, tops BrowseComp & SWE-Bench, and collapses the gap between open and closed AI.
“AI Quill” Publication 200 Subscriptions 20% Discount Offer Link.
Last week, Kimi officially announced the release and open-sourcing of the K2 Thinking model. This newcomer achieved SOTA results on Agentic benchmarks such as Humanity’s Last Exam, BrowseComp, and SEAL-0, natively supports INT4 inference, and can complete up to 300 rounds of tool calls without human intervention...
At first glance, this might seem like just another breakthrough in domestic AI model capabilities. However, a quiet revolution in how users interact with AI has also begun.
Let’s review how we used to interact with AI. There were several approaches:
Command-style: “Help me look up XX” or “Summarize this document.”
Template-style:
You are a senior XXX, my background is XXX, I need you to help me XXX. Copy a bunch of formatted prompts and fill in the task information.Expert-style: Write in JSON/DSL format, hard-coding roles, processes, and formats, like this: