Kimi K2: the new SOTA open source AI model is Chinese!

Published on

Hot day in the world of open source AI models! A new model came out that, within 25 minutes of being published, already had more than 12 thousand downloads!

Moonshot AI, a Chinese company backed by other big players like Tencent and Alibaba, just launched “Kimi K2”, a new Mixture-of-Experts (MoE) model with impressive specs: 1 TRILLION total parameters and 32 billion activated parameters.

On benchmarks, it also does really well, becoming the new state of the art among open source models in several categories, especially in programming and math/logical reasoning, going head to head with giants like DeepSeek-V3 and Qwen. In some tests, it even surpasses proprietary models like Claude Sonnet 4 and Gemini 2.5 Flash.

To be able to train a model this size (on 15.5 trillion tokens, a number similar to the training estimates for GPT-4 and 4o) in a stable way, they had to develop a new optimizer called MuonClip, which solved instability problems that other optimizers like AdamW had at this scale.

The model was published on Hugging Face in two versions (Base and Instruct) and with a Modified MIT license, which is pretty permissive.

You’ll need a LOT of GPU to run a model this size, but either way, excellent news for the open source community!