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.

Table of Contents

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
4. TRI’s View

<Total Pages: 26>