2.8 Trillion! China's Moonshot AI Releases the World's Largest Open-Source Model by Parameters

en.Wedoany.com Reported - Beijing-based Moonshot AI officially released its next-generation large model, Kimi K3, on July 16, 2026. With a total parameter scale of 2.8 trillion, it is currently the world's largest open-source artificial intelligence model. On the eve of its release, the 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance opened in Shanghai.

Kimi K3 is built on the KDA Hybrid Linear Attention Mechanism (Kimi Delta Attention) and attention residual technology, natively supporting visual understanding with a context window of 1 million tokens. The model is optimized for complex task scenarios such as software engineering, knowledge work, deep research, and multimodal understanding, aiming to enhance the processing capabilities of large models for complex tasks.

In terms of performance, Kimi K3's comprehensive intelligence level approaches that of leading closed-source models globally. In the Frontend Code Arena evaluation, Kimi K3 scored 1679 points, surpassing Claude Fable 5 to rank first. Among the models tested, its comprehensive intelligence level is second only to the two leading closed-source models, Claude Fable 5 and GPT-5.6 Sol. Moonshot AI stated that Kimi K3's overall performance in evaluations still lags behind the strongest closed-source models, Claude Fable 5 and GPT-5.6 Sol.

In terms of application capabilities, Kimi K3 can not only complete explicit code generation and modification tasks but also engage in open-ended creation and complex knowledge production. In digital content generation scenarios, the model can combine 3D reasoning, programming, and visual capabilities to transform concepts, images, or videos into interactive experiences. In scenarios such as financial research and industry analysis, it can organize and analyze company documents, industry data, and long-term industry materials to generate readable, searchable, and verifiable in-depth reports.

Moonshot AI stated that the model's complete weights will be released soon, and details on the model architecture, training, and evaluation will be published gradually along with the model's technical report. Liu Tieyan, Dean of the Beijing Zhongguancun Academy, noted that a group of Chinese open-source large models are moving from "individual debuts" to "collective breakthroughs," offering new solutions and pathways for the global development of artificial intelligence.