Kimi K2 by Moonshot AI is a new 1T parameter open-source MoE model (32B active). It delivers state-of-the-art performance on coding, reasoning, and agentic tasks. Base and Instruct models are now available.

Hi everyone!

A 1T parameter open-source base model... with benchmark results this impressive. I really hope the delay of OpenAI's next open model isn't because they saw Kimi K2 coming🤔

But regardless, I know the Moonshot team has been putting in the work for this moment, and they absolutely deserve it. They've given the world an incredible open-source model. As far as I know, many dev communities are already incredibly impressed with its agentic coding capabilities.

A heads-up though: given this is a 1T parameter model, running it locally on consumer hardware will be tough for most. The easiest way to get started is by Kimi's official website or API.

Running Kimi K2 locally needs very powerful hardware, like top-tier GPUs or expensive computers, which most people don’t have. How do you plan to make Kimi K2 more accessible for users with regular hardware?

@vouchy I don't think this is the point. It's more that you have access to it and have full confidence that this is the model you're using, instead of self-hosting a 1T parameter (32b active) LLM on your own.

That's why we have services like OpenRouter, and even the official API from Moonshot.

Kimi K2

The 1T parameter open model for agentic intelligence