Kimi K3 Dominates Global Open-Source Models: Core Technologies and the New US-China AI Race

Description

On July 27, 2026, the Chinese AI startup Moonshot AI officially open-sourced the weights for its flagship AI model, Kimi K3. Boasting a total parameter count of 2.8 trillion (2.8T), Kimi K3 currently stands as the world’s largest open-source model, establishing a new benchmark for frontier AI models in China. Shortly after, DeepSeek released DeepSeek V4 Flash on July 31, and Alibaba Cloud announced Qwen 3.8 Max on August 3. These rapid developments have fueled market discussions regarding a potential “second DeepSeek moment” and the intensifying US-China AI rivalry.

This report provides an in-depth analysis of the following: (1) an overview of Kimi K3, Qwen 3.8 Max, and DeepSeek V4 Flash; (2) a dissection of Kimi K3’s core technologies; and (3) a comparative analysis of AI models developed in China and the United States. This report aims to evaluate mainstream Chinese AI models, elucidate the underlying technical principles and trends of Kimi K3, and assess the competitive landscape between Chinese and American AI capabilities.

Key Highlights

Moonshot AI open-sourced Kimi K3, now the world's largest open-source AI model, setting a new benchmark for China's frontier AI development.
DeepSeek and Alibaba Cloud quickly followed with new model releases, sparking talk of a "second DeepSeek moment" and sharpening US-China AI rivalry.
The report examines these leading Chinese models, unpacks Kimi K3's core technology, and compares the competitive landscape between Chinese and US AI capabilities.

Product Code: TRi-211

1. Overview of Kimi K3, Qwen 3.8 Max, and DeepSeek V4 Flash

Figure 1: Example of Kimi Agent in Use-Simulating the Gargantua Black Hole
Figure 2: Comparison of Total Parameters Across Mainstream Open-Source AI Models
Figure 3: Comparison of Active Parameters Across Mainstream Open-Source MoE AI Models
Table 1: Evolution of Specifications for Moonshot AI’s Kimi Models
Table 2: Evolution of Specifications for DeepSeek’s AI Models
Table 3: Evolution of Specifications for Alibaba’s Qwen Models
Table 4: Comparison of Kimi K3, Qwen 3.8 Max, and DeepSeek V4 Flash

2. Analysis of the Core Technologies of Kimi K3

Figure 4: Architecture of Kimi K3
Figure 5: Architectures of Kimi-VL and Kimi-VL-Thinking
Table 5: Comparison of First-Generation MoonViT and MoonViT-V2
Figure 6: Schematic Diagram of Softmax Attention vs. Linear Attention Principles
Figure 7: Architecture of Kimi Linear Hybrid KDA-MLA
Figure 8: Architecture of LatentMoE
Figure 9: SiTU-GLU Boundary Setting
Figure 10: Schematic Diagram of Quantile Balancing Mechanics
Figure 11: Architecture of Kimi AttnRes
Figure 12: Role of Optimizers in AI Training (Using Decoder-Only Transformer as Example)
Figure 13: Efficacy of MuonClip
Table 6: Technical Comparison between Kimi K3 and DeepSeek V4

3. Comparison between Chinese and US AI Models

Figure 14: Trend of Intelligence Index among Mainstream AI Models (November 2022 to July 2026)
Figure 15: Trend of Intelligence Index among Most Advanced Chinese and US AI Models (November 2022 to July 2026)
Figure 16: Percentage Trend of Intelligence Index Gap between Most Advanced Chinese and US AI Models (October 2024 to July 2026)
Table 7: Comparison of Performance and API Pricing between Kimi K3, Claude Opus 5, and GPT 5.6 Sol
Table 8: Hypothetical Estimation of Major Costs
Figure 17: Cost Comparison between Using API and Self-Deploying Kimi K3