Moonshot AI releases weights of popular Kimi-K3 — nearly on par with market leaders but much simpler to run

Moonshot AI has released the weights of its Kimi-K3 model for free, allowing anyone with sufficient compute to run it locally. The model significantly outperforms previous generations of OpenAI GPT and Anthropic Claude, and only slightly lags behind the latest Sol and Fable models. Kimi-K3 is 2-3 times cheaper to operate than American counterparts, and with cache hits on coding tasks, costs drop to $0.30 per million tokens versus $3 standard input pricing. The efficiency likely comes from using MXFP4 for weight storage and MXFP8 for activation, both low-precision formats that save memory. Of its 2.8 trillion parameters, only 104.2 billion are active simultaneously. The model is designed to run on Nvidia H20 accelerators, which lack native MX floating-point support, but may run even more efficiently on Blackwell chips. Kimi-K3 replaces key-value cache with a fixed-size state handler called Kimi Delta Attention, and uses MoE with only 16 out of 896 experts activated. However, only the weights are open; the training process and dataset remain proprietary.