Who Built Kimi K3? Inside the Elite Team Driving a $70 B Valuation

Open‑source release of Kimi K3

On 27 July 2024 Moonshot AI released the full 2.8‑trillion‑parameter Kimi K3 model and its technical report, listing 401 contributors.

Key technical contributions

Mooncake KV‑Cache architecture – a decoupled inference system that separates the pre‑fill and decode stages, aggregates idle CPU, memory and SSD resources into a distributed cache pool. The design improves request‑handling capacity by >75 % and reduces token‑level latency by up to 94 % for long‑context workloads. Implemented in the open‑source repository kvcache‑ai/Mooncake (GitHub, >6 000 stars) and awarded the FAST 2025 Best Paper prize.

MoBA (Mixture‑of‑Block‑Attention) – block‑level sparsity applied to attention layers, achieving 95 % sparsity with no loss in accuracy. Benchmarks show a 6.5× speed‑up for million‑token sequences and a 16× speed‑up for ten‑million‑token sequences.

Muon optimizer – replaces AdamW with a matrix‑orthogonalization scheme, roughly halving the training compute required for trillion‑parameter models. Described in the paper “Muon is Scalable for LLM Training” and used in the open‑source training framework Moonlight.

Attention Residuals – redesign of residual connections that mitigates gradient decay in very deep networks, yielding a 1.25× increase in decoding throughput for 7 B models.

Kimi Linear – a linear‑attention architecture that reduces KV‑Cache size by up to 75 % and raises decoding throughput six‑fold for million‑token contexts.

Performance results

The K3 technical report reports inference cost at 38 % of Claude Fable 5, a request‑handling capacity increase of >75 %, and long‑context efficiency improvements ranging from 59 % to 498 % depending on token length. Daily token throughput exceeds 1 000 billion tokens across thousands of production nodes.

Selected contributor technical backgrounds

Zhou Xinyu – co‑founder and algorithm lead; previously at Megvii where he authored the lightweight ShuffleNet architecture. Leads large‑model algorithm production and hybrid architecture optimization (KDA‑MLA).

Wu Yuxin – co‑founder; primary author of Detectron2 and co‑author of the Group Normalization paper (ECCV 2018 Best Paper Honorable Mention). Designed the KDA‑MLA hybrid architecture that reduces pre‑training and RL costs.

Zhang Yuta – CTO; built the AMiner big‑data platform and long‑context knowledge‑graph techniques that underpin Kimi’s ultra‑long‑text capabilities.

He Weiran – inference system lead; co‑author of MoBA, responsible for the 6.5× and 16× speed‑ups at 95 % sparsity.

Du Yulun – pre‑training and scaling lead; co‑author of Attention Residuals, MoBA and the Muon optimizer, which halves training compute.

Chen Guangyu – contributed to Attention Residuals, achieving a 1.25× decoding throughput improvement for 7 B models.

Qin Ruoyu – PhD student at Tsinghua MADSys; designed the Mooncake KV‑Cache‑centric serving platform, separating pre‑fill and decode stages, increasing request capacity >75 % and reducing token latency up to 94 %.

Liu Shaowei – infrastructure lead; pioneered extreme INT4 quantization and the KTransformers project that enables trillion‑parameter MoE inference across heterogeneous clusters.

Open‑source artifacts

Technical report: https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf<br/> Mooncake repository: https://github.com/kvcache-ai/Mooncake (stars > 6000)<br/> Moonlight training framework and Muon optimizer: https://github.com/MoonshotAI/Moonlight