Kimi K3 makes frontier AI more sovereign, not simple

Subject: Kimi K3 makes frontier AI more sovereign, not simple

Preview: Open weights, 2.8 trillion parameters and a one-million-token context window.

Moonshot AI released the weights of Kimi K3, a native multimodal model with 2.8 trillion parameters and a one-million-token context window. It activates 16 of 896 experts for each token, reducing the amount of computation used at each step.

The company published the model weights, a technical report and parts of the deployment stack. It positions K3 for long-horizon coding, research, agentic work and multimodal analysis.

Organisations gain more options for hosting, data control, customisation and supplier continuity. In regulated sectors, that freedom can matter as much as a top benchmark score.

Open does not mean small or cheap. A model of this scale needs substantial infrastructure, optimisation expertise and licence governance. Kimi's own report also says the strongest proprietary systems still lead on overall performance.

Compare three scenarios: external API, managed hosting and self-deployment. Measure total cost, latency, confidentiality, controls, available skills and quality on your own tasks.

Open weights expand your choices; they do not remove the cost of operating the model.