The next-generation large model Kim K3 from Moonshot AI will be released this month, with an expected parameter scale of 2.5 trillion, surpassing DeepSeek V4 Pro and Baidu's Wenxin 5.0, and poised to become the largest Chinese large model by parameter size. The model is expected to deeply integrate a 1M ultra-long context window with multimodal processing capabilities. Meanwhile, Google's Gemini 3.5 Pro has been delayed until July, and Elon Musk’s Grok v9 is expected to be completed by the end of July. Competition among large models is intensifying, and breakthroughs in parameter scale and multimodal capabilities by Chinese models will provide developers and enterprises with more high-performance tool options.Author and source: AIBase
The race in the large model field begins in the second half of 2026.
The race in the large model field has entered a heated phase right from the start of the second half of 2026. In addition to the highly anticipated official release of DeepSeek V4, Kimi K3, the next-generation large model from AI unicorn Moonshot AI, has also been officially confirmed for release this month.
Although the official release date and detailed technical documentation for K3 have not yet been disclosed, insiders have indicated that this new model represents a significant leap in performance. It is expected to feature a parameter scale of up to 2.5 trillion, substantially surpassing DeepSeek V4 Pro’s 1.6 trillion and surpassing Baidu’s previously released ERNIE 5.0 (2.4 trillion), positioning it as the largest parameter-scale large model among China’s domestic AI models.
In terms of functionality, Kimi K3 is expected to deeply integrate a 1M ultra-long context window with multimodal processing capabilities. This combination of “massive parameters + comprehensive multimodality” is regarded by the industry as a key advantage for Chinese models to compete at the global top tier. Previously, Moonshot AI launched Kimi K2.7 Code and its high-speed version in June, both specialized for AI programming, maintaining a high-frequency iteration pace of “one update per month.” The release of K3 will undoubtedly further solidify its first-mover advantage in the technical ecosystem.
Meanwhile, the large model market in July has been fiercely competitive. On the international front, Google’s Gemini 3.5 Pro, originally scheduled for a June release, has been delayed until this month; Musk’s Grok series has also been highly active, with its deep integration into the AI coding tool Cursor, and the Grok v9 model, targeting Opus-level performance, is expected to be completed by late July and officially launched in August.
As model vendors continue to advance their underlying algorithms and commercial applications, the artificial intelligence market landscape in the second half of the year will become increasingly clear. For developers and enterprise users, the dual breakthroughs of domestic large models in parameter scale and multimodal capabilities mean that more high-performance, cost-effective professional tools will be available in the future. The industry widely anticipates that Kimi K3 will deliver exceptional performance, elevating the domestic large model ecosystem to new heights of intelligence.
