AI industry update

研报英文原文证据摘录

AI industry update

Global Markets Research

22 July 2026AI industry update

EQUITY: JAPAN ELECTRONIC PARTS

Research Analysts

Moonshot AI releases Kimi K3 Japan electronic parts

Manabu Akizuki - NSC

Takes first step in post-transformer world with open-source, large-scale manabu.akizuki@nomura.com

coding model +81 3 6703 1185

Kimi K3: Groundbreaking innovation with hybrid architecture featuring Transformer

and recurrent model

Moonshot AI's Kimi K3, released on 16 July, has become a hot topic as it has coding

performance on par with that of cutting-edge AI lab models in the US. On top of this, its

innovative hybrid model combines Transformer (multi-head latent attention; MLA) with a

recurrent model (Kimi Delta Attention; KDA), and we were surprised by this. In terms of

implications for the technology industry, we think the emergence of Kimi K3 provides fresh

evidence of China's strong ability to develop AI models, as the cutting-edge agentic

coding model has shaken up the dominant position of the US's closed models.

Achieves long-term reasoning by substantially reducing KV cache with a recurrent

model using KDA

Kimi K3 is a mixture of experts (MoE) model with 2.8trn parameters. It is made up of 896

experts, and when carrying out inference, the router selects and activates 16 experts per

token. The maximum context length is 1mn tokens, and the model is specialized for long-

term agentic coding.

As noted above, the core of the model is a hybrid architecture that combines Transformer

with a recurrent model. Based on the company's past announcements, we surmise that

the attention architecture consists of four layers as one unit: three KDA layers and one

MLA layer. While the Transformer model holds the keys and values of all tokens in the