Japan IT services: July AI model rankings: Rise of Chinese open-weight models, quantization, and AI model commoditization

研报英文原文证据摘录

Japan IT services: July AI model rankings: Rise of Chinese open-weight models, quantization, and AI model commoditization

Viewpoint |

21 Jul 2026 04:27:57 ET │ 12 pages

Japan IT services

July AI model rankings: Rise of Chinese open-weight models,

quantization, and AI model commoditization

Moriya Koketsu AC

CITI'S TAKE +81-3-6776-8304

The AI model development race is heating up and becoming increasingly moriya.koketsu@citi.com

multifaceted. In the benchmark race, Moonshot AI's Kimi K3 from China can Masahiro Shibanoperform as well as earlier frontier models. Meanwhile, Meta's Muse Spark 1.1

and xAI's Grok 4.5 leverage Mixture of Experts (MoE) technology to deliver +81-3-6776-4641

top-tier AI performance at lower costs. Furthermore, PrismML's Bonsai 27B masahiro.shibano@citi.com

compresses a 27bn-parameter model using quantization technology,

enabling it to run on an iPhone. The series of AI advances in July marked a

turning point in AI development. The performance gap between US and

Chinese models has narrowed significantly, and AI model commoditization

has progressed to levels that raise concerns about frontier AI companies'

margins.

AI model leaderboard —The biggest shock in July's AI model rankings was the

emergence of Moonshot AI's Kimi K3, which ranked 1st in Arena Ai’s front-end coding

rankings. Its output price of $15 per million tokens is well below output prices at top-

flight competitors. Kimi K3's full capabilities will be known when it is released as an

open-source model on July 27. If a frontier-class capability open-weight model is

released, it will promote AI usage while significantly increasing security concerns.

MoE —The emergence of Meta's Muse Spark 1.1 and xAI's Grok 4.5 represents

another important development.