Ruoyu Qin et al. at Moonshot AI and Tsinghua published Mooncake, introducing a KVCache-centric disaggregated serving architecture that separates prefill and decoding clusters to handle Kimi's multi-hundred-thousand-token prompts under overload.
Moonshot AI & Kimi
Also known as: Moonshot AI · Kimi · Kimi k1.5 · Kimi K2 · Mooncake
Plain English
The Chinese AI research lab founded by Transformer-XL/XLNet co-author Yang Zhilin—builder of the long-context Kimi assistant, the Mooncake disaggregated serving architecture, and the 1-trillion-parameter open-weight Kimi K2 MoE model.
Technical Overview
Moonshot AI pioneered KVCache-centric disaggregated serving (Mooncake), multimodal long-to-short reinforcement learning (Kimi k1.5), and 1T-parameter MoE pretraining stabilized via the MuonClip (qk-clip) optimizer (Kimi K2, 1T total / 32B active parameters) tailored for long-horizon agentic tool use.
Timeline (4)
2024
Paper2025
PaperMoonshot AI published Kimi k1.5: Scaling Reinforcement Learning with LLMs, matching o1-class multimodal reasoning without Monte Carlo tree search or value functions by combining 128k-context RL scaling with long-to-short chain-of-thought compression.
- Release
Moonshot AI open-sourced Kimi K2, a 1-trillion-parameter MoE (32B active parameters, 384 experts) pretrained on 15.5T tokens with the MuonClip optimizer (
qk-clipattention logit bounding) for zero loss spikes and post-trained on large-scale synthetic MCP and agentic coding trajectories. - Release
Moonshot AI released Kimi K2 Thinking, applying native INT4 Quantization-Aware Training (QAT) during post-training and enabling interleaved chain-of-thought reasoning across 200–300 consecutive tool calls.
See also (11)
Open full graphMoonshot AI & Kimi