Domestic AI Large Model Development Accelerates: The Two-Trillion Parameter Club Takes Shape

Recently, China's large models have achieved successive breakthroughs. On July 17, Moonshadow released the Kimi K3 model, which boasts 2.8 trillion parameters and a 1 million token context window, making it the world's largest open-source weight model by parameter count. In the intelligence scoring by independent AI model evaluation agency Artificial Analysis, K3 ranked third, narrowly trailing Claude Fable 5 from Anthropic and GPT-5.6 Sol from OpenAI. On July 19, Alibaba's Tongyi Qianwen launched the preview version of its Qwen3.8-Max model and announced that Qwen3.8 would soon be released and open-sourced. This model has a parameter scale of 2.4 trillion and is described as "possibly the most powerful model aside from Fable 5." With trillion-level parameter scales and performance comparable to top-tier closed-source models, domestic AI now has the confidence to raise prices. The API call fees for the K3 model have significantly increased compared to its predecessor, K2.6, making it the most expensive domestic large model currently available. Industry insiders view this as an important signal, indicating that domestic large models are transitioning from cost-effective alternatives to technological competitors. (Securities Times)