Kimi K3 emerges spectacularly; Morgan Stanley: China's leading large models have caught up across the board, simultaneously achieving breakthroughs in scale, performance, and pricing.

Kimi K3 emerges spectacularly; Morgan Stanley: China's leading large models have caught up across the board, simultaneously achieving breakthroughs in scale, performance, and pricing.

Moonshot released its flagship model K3 on July 17, ranking among the top three in global benchmark tests and attracting widespread market attention. Morgan Stanley characterized the launch of K3 as a symbolic sign of China’s large model industry fully catching up to U.S. leaders—a catch-up that is occurring simultaneously in the dimensions of model scale, performance, and API pricing.

According to Windchaser Trading Desk, Morgan Stanley pointed out in its latest report that K3 ranked third in Artificial Analysis’s Global Intelligence Index, just behind two top American large models; it also took first place globally in Arena.ai’s programming leaderboard. Meanwhile, K3’s API pricing is the highest among major Chinese large models, further narrowing the price gap with top U.S. products and having a positive impact on overall market pricing structures.

Morgan Stanley analysts Gary Yu and Lydia Lin emphasized in their report that K3’s emergence is not an overnight miracle, but the result of long-term accumulated progress in China’s AI model industry. The report predicts that more Chinese large models will continuously break through in parameter scale, pricing level, and performance, with strong global competitiveness.

K3 Core Parameters: 2.8 Trillion Parameters, Largest Open-Source Chinese Model to Date

K3 is an open-weight large language model focused on long-term tasks, knowledge work, and reasoning capabilities, with a total of 2.8 trillion (2.8T) parameters, making it currently the largest open-source Chinese model. Moonshot plans to fully release all weights by July 27.

Technically, K3 adopts two original algorithmic innovations—Kimi Delta Attention and Attention Residual—aiming to improve computational efficiency. The context window is 1 million tokens, on par with mainstream Chinese models such as Alibaba Qwen3.7-Max, MiniMax M3, and Zhipu GLM-5.2.

According to Morgan Stanley’s comparative data, front-running Chinese large models currently show significant differences in parameter scale: MiniMax M3 has 428 billion parameters, DeepSeek V4-Pro has 1.6 trillion, Zhipu GLM-5.2 has 754 billion, and K3 leads with 2.8 trillion parameters.

Pricing Strategy: High Prices Reshape Market Dynamics

In terms of API pricing, K3’s output price is 100 RMB per million tokens, far higher than DeepSeek V4-Pro and Xiaomi MiMoV2 at 6 RMB, and also higher than Zhipu GLM-5.2 at 28 RMB and Alibaba Qwen3.7-Max at 36 RMB.

The input price (uncached) is 20 RMB per million tokens, also the highest among major Chinese large models.

Morgan Stanley believes that K3’s high pricing strategy has a positive effect on the large model market dynamics. Previously, China’s large model market had fierce price competition; K3’s pricing direction helps guide the industry toward a more sustainable business model and narrows the pricing gap between Chinese and American large models.

Performance: Ranked First Globally in Programming, Chasing Top U.S. Products Overall

In performance benchmarks, K3 scored 57 in Artificial Analysis’s Intelligence Index, significantly ahead of mainstream Chinese models at the time—MiniMax M3 and DeepSeek V4-Pro both scored 44, Alibaba Qwen3.7-Max scored 46, and Zhipu GLM-5.2 scored 51.

In programming capabilities, K3 ranks first globally in Arena.ai’s programming leaderboard, surpassing all American competitors.

Morgan Stanley’s report notes that K3 has received positive feedback worldwide, marking substantial progress by Chinese large models in core capabilities, closing in on U.S. leaders.

Morgan Stanley’s Assessment: Not an Overnight Miracle, But a Concentrated Result of Industrial Accumulation

Morgan Stanley explicitly states that K3 should not be seen as a sudden technical breakthrough, but rather understood within the continuous evolution of China’s AI model industry. The report indicates that Zhipu GLM-5.2, released in June this year, already demonstrated excellent performance, and K3’s debut is a continuation and enhancement of this progress.

The report also points out that the leap in K3’s parameter scale shows that the scaling law remains effective at this stage, which is important for assessing the tech roadmap of China’s large model industry.

Morgan Stanley predicts that China will continue to see the emergence of larger, higher-priced, stronger-performing large models, forming a powerful competitive position in the global market.

 

 

 

~~~~~~~~~~~~~~~~~~~~~~~~

The above highlights are from Windchaser Trading Desk.

For more in-depth interpretation, including real-time analysis and frontline research, please join [Windchaser Trading Desk · Annual Membership]

Risk Warning and DisclaimerThe market has risks, and investment needs caution. This article does not constitute personal investment advice, nor does it consider the individual investment objectives, financial situations, or needs of specific users. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their particular circumstances. Investment is at their own risk.