Moonshot AI launches Kimi K3 reasoning model on Amazon Bedrock

Moonshot AI has launched its Kimi K3 reasoning model on the Amazon Bedrock cloud platform. Available in all AWS regions, the model is designed for enterprise use and features a context window of one million tokens.

Kimi K3 utilizes a Mixture-of-Experts (MoE) architecture, consisting of 2.8 trillion total parameters and 896 routing experts. During a forward pass, approximately 104 billion parameters are activated. The model also includes native image processing capabilities and is noted as the first open-weights model of its scale available on the Amazon platform.

To improve efficiency, the system implements prompt caching starting at 1,024 tokens, with a minimum cache duration of 30 minutes. Moonshot AI claims the model achieves 2.5 times the scaling efficiency of its predecessor, Kimi K2, though this has not been independently verified.