Moonshot AI

Chinese AI company (Kimi). Developers of Mooncake (KVCache-as-resource serving system) and the Kimi Linear architecture with hybrid attention. Collaborators with vLLM, SGLang, and Dynamo on open-source serving.

Key Contributions

• Mooncake: First-class KVCache resource management for PD disaggregation
• Kimi Delta Attention (KDA): Hybrid attention reducing KVCache by ~10x
• PrfaaS: Cross-datacenter prefill-decode serving architecture
• Partnership with Tsinghua University on serving systems research

Connections

• PrfaaS — cross-datacenter serving paper
• PD Disaggregation — the serving paradigm they’re advancing