Kimi K2 Thinking - The 1T Parameters Open-Source Thinking Model - SOTA on HLE
🔹 SOTA on HLE (44.9%) and BrowseComp (60.2%)
🔹 Executes up to 200 – 300 sequential tool calls without human interference
🔹 Excels in reasoning, agentic search, and coding
🔹 256K context window
Replies
Kimi AI - Now with K2.6
👋 Hello from Kimi Team!
Introducing Kimi K2 Thinking: 1T Open-Source Reasoning Model. SOTA with 44.9% HLE, 60.2% BrowseComp. (not just open-source SOTA)
> Trillion-param MoE, trained for $4.6M, 4x cheaper than peers.
> INT4 inference: 4-bit quantized, <1.2s latency @ 256K context.
> Full step-by-step reasoning, 200+ tool calls, self-correction (GPT-5 level), fully open (MIT), OpenAI-compatible API, weights live on huggingface today, agentic mode next week.
We're thrilled to ship a SOTA model that's fully open. Can't wait to see what you all build! :)
Flowtica Scribe
Hi everyone!
The K2 model from July was already strong, but it wasn't a "Thinking" model.
Kimi K2 Thinking is a new generation thinking agent model, built on Moonshot's "model as agent" philosophy. The key difference is that it natively understands how to "think while using tools."
It can build a real, functional Word editor:
And it can also create a world of complex, gorgeous voxel art:
@zaczuo Would be cool if it could team up with other agents to tackle bigger stuff
Secure MCP Framework by Arcade.dev
Woah, quantized INT4 inference is a big deal here! Congrats on the launch!
Swytchcode
Super responses! But why did you focus so much on slides?
It’s a bit inconvenient that you can’t use it without registering. You even need to sign up for the first search. But the project is interesting anyway!