On July 17, 2026, in the early hours, Moonshot AI unveiled its next-generation large model, Kimi K3, boasting a parameter scale of 2.8 trillion—making it the world’s first open-source model at the 3-trillion parameter level. On the day of its release, the Hong Kong-listed "first large model stock" Zhipu (02513.HK) plunged 28.49%, wiping out over HKD 200 billion in market capitalization in a single day. As of July 20, Zhipu’s share price had dropped further to HKD 890.5, with a market cap of HKD 414.637 billion.
That same week, DeepSeek reportedly completed a $7.4 billion funding round, setting a new record for the largest private placement in China’s AI sector, with the company valued at over $50 billion. Just a month prior, Zhipu had launched and open-sourced GLM-5.2, a model focused on long-range tasks.
Three major events concentrated within less than a month signal a new phase in China’s AI large model industry. The parameter scale race is far from over, but the logic of valuation in capital markets is shifting. This article analyzes the competitive landscape of Chinese AI large models post-Kimi K3 from four perspectives: the relationship between parameters and valuation, the impact of open-source models, differences in commercialization strategies, and long-term moats.
Why Parameter Scale No Longer Determines Valuation
Kimi K3’s parameter scale reaches 2.8 trillion, far surpassing Zhipu GLM-5.2 and the DeepSeek-V4 series. Moonshot AI likens parameters to neural connections in the human brain, with nearly 3 trillion parameters allowing the model to store more knowledge and patterns.
Yet, capital markets responded in divergent ways. Zhipu’s stock plummeted 28.49% on Kimi K3’s launch day, despite Zhipu announcing a HKD 31.4 billion placement just days earlier on July 9—a record for single fundraising by a Hong Kong tech company in 2026. This data reveals a key shift: parameter scale alone is no longer a sufficient condition for valuation.
Industry data shows that actual usage of Chinese large models is rising rapidly. According to OpenRouter estimates, last week (June 29–July 5), global AI large model calls totaled 46.7 trillion tokens, with Chinese AI large models accounting for 23.45 trillion tokens—a week-on-week increase of 15.01%—maintaining the lead over the US for ten consecutive weeks. The top six global models by usage are all Chinese, with DeepSeek-V4-Flash holding the top spot for seven weeks straight, reaching 5.34 trillion tokens per week.
The surge in usage reflects real integration into enterprise workflows. As industry reports point out, the Chinese AI large model sector in the first half of 2026 has moved from pure technical competition to a full-scale "commercial deployment battle," validating application prospects and beginning to clear industry bubbles.
Jefferies, in its research following Kimi K3’s launch, noted that the model has entered the global frontier in coding, agent, and multimodal tasks. However, its significance lies not just in improved capabilities, but in confirming a shift in China’s AI industry’s competitive logic. Whereas the market once focused on "large model capability gaps," the current competition is shifting toward "unit intelligence cost" and "token consumption scale."
How Open-Source Models Are Changing Industry Competition
Kimi K3’s open-source strategy is one of the most noteworthy variables in this release. As the world’s first open-source model at the 3-trillion parameter level, K3’s full weights are expected to be released by July 27 at the latest. Prior to this, Zhipu had open-sourced GLM-5.2 on June 17, which scored 51 points on Artificial Analysis’s latest global large model intelligence index, ranking first among open-source models.
The competitive logic for open-source models fundamentally differs from that of closed-source models. Closed-source models build their advantages on proprietary technology barriers and API pricing power, while open-source models derive value from ecosystem development and developer stickiness. Kimi K3 uses an MoE (Mixture of Experts) architecture, reducing actual compute requirements via expert routing. Its API pricing is $3 per million tokens for input and $15 per million tokens for output. DeepSeek’s V4 Flash model delivers standard intelligent tasks at a weighted average cost of just $0.02, while Anthropic’s comparable model costs $2.75.
This cost difference is reshaping the global AI supplier landscape. San Francisco-based Lindy AI switched from Anthropic to DeepSeek after six weeks, cutting costs to just 10% of the original and saving millions annually. Airbnb, Pinterest, and even Microsoft are now using Chinese open-source AI models for routine coding, automation, and daily tasks. At times this year, Chinese models have captured over 30% market share on the AI aggregation platform OpenRouter.
Another major impact of open-source models is the lowered commercialization threshold for AI applications. As Chinese models approach top overseas models in performance while maintaining low inference costs, token usage growth is spreading to cloud computing, AI chips, and infrastructure. SenseCore, the large-scale device from SenseTime, is shifting its compute resources from training to inference, aiming to prove itself not just as a large model company but as a full-stack AI service provider.
Comparing Commercialization Paths of AI Companies
The three leading companies each follow distinct commercialization strategies.
Zhipu pursues a "public listing + large-scale financing" capital-driven path. Listed on the Hong Kong Stock Exchange on January 8, 2026, its share price has surged over 1,400%. The July 9 placement plan proposes issuing up to 19.78 million new H shares at HKD 1,588 each, raising about HKD 31.41 billion. As of June 30, over 93% of IPO proceeds had already been used. Funds are earmarked for three main areas: foundational model R&D and compute infrastructure, commercialization and strategic industry investments, and capital structure optimization. Zhipu is also advancing a listing on the STAR Market in mainland China, aiming to raise RMB 15 billion. Total fundraising has reached HKD 36.4 billion (about RMB 31.5 billion).
DeepSeek follows a "private placement + self-built infrastructure" path. This Hangzhou AI lab, founded just three years ago, trained high-performance models with far less compute than OpenAI or Anthropic. At the end of May 2026, it completed its first funding round of about $7 billion, with a post-money valuation of $52 billion. Just a month later, it reportedly launched a second round, with a pre-money valuation of $71 billion. DeepSeek is building a gigawatt-level compute campus, with investments in self-built data centers possibly reaching hundreds of billions of RMB. The company is also developing its own AI inference chips. DeepSeek plans to IPO in 2027.
Moonshot AI (Kimi) takes a "product-driven + open-source ecosystem" approach. After the K3 model launch, President Zhang Yutong disclosed that the company’s ARR (annual recurring revenue) grew severalfold. Yet just three days after release, Moonshot AI announced it would pause new subscriptions for consumer users, dedicating all existing compute resources to current subscribers. Compute shortages are both a challenge and a signal—demand is real, but supply bottlenecks are constraining growth. Moonshot AI says it is rapidly expanding compute capacity. On the commercialization front, the company could go public in as little as six months.
Each path has its strengths and weaknesses. The capital-driven route offers abundant funding and strong risk resistance but requires ongoing proof of growth logic to capital markets. Private placement is more flexible but faces valuation pressure. Product-driven strategies foster the strongest user stickiness, but are most constrained by compute and funding.
Who Has a Long-Term Moat?
Assessing long-term moats requires looking at three dimensions: technological barriers, cost advantages, and ecosystem stickiness.
Technological barriers: Kimi K3 topped the Frontend Code Arena leaderboard with a score of 1,679, surpassing Claude Fable 5’s 1,631. However, overall evaluations show K3 still trails closed-source models like Claude Fable 5 and GPT-5.6 Sol. Chinese models currently rank around fifth globally, and have yet to challenge the dominance of overseas giants. The technology gap is narrowing, but hasn’t disappeared.
Cost advantage is the most prominent moat for Chinese AI companies. DeepSeek’s ultra-low pricing strategy has attracted developers worldwide. Jefferies notes that as Chinese models approach top overseas models in performance while maintaining lower inference costs, the commercialization threshold for AI applications may drop further. This cost advantage stems not only from architectural innovation but also from China’s lower electricity and data center infrastructure costs.
Ecosystem stickiness may be the hardest moat to replicate. Open-source models build ecosystems through developer communities, while closed-source models lock in enterprise clients via API services. DeepSeek-V4-Flash has led in usage for seven consecutive weeks, indicating strong developer ecosystem traction. Zhipu is accumulating industry know-how by focusing on verticals like government, finance, and manufacturing. If Kimi K3’s open-source strategy attracts enough developers to build applications on it, its ecosystem advantage will gradually emerge.
In summary, no company currently possesses an absolute long-term moat. The technology gap is shrinking but persists, cost advantages can be imitated, and ecosystem stickiness takes time to build. The real moat may not lie in any single dimension, but in the ability to create a positive cycle among technology, cost, and ecosystem—technological breakthroughs lower costs, low costs attract more developers, more developers drive ecosystem prosperity, and ecosystem prosperity feeds back into technological progress.
Conclusion
The launch of Kimi K3 is a landmark event, but it signals not the victory of any single company, but a turning point for the industry. China’s AI large model sector is moving from a "parameter scale race" to a commercialization competition focused on "lower costs, higher usage, and stronger ecosystems."
As of July 20, 2026 (Beijing time), in the cryptocurrency market, Bitcoin (BTC) found support near $64,000 and rebounded steadily, reaching an intraday high of about $65,000. Gate quotes BTC at around $64,747. The market is being pulled by two forces: Middle Eastern geopolitical tensions are driving up oil prices, while investors are still digesting the shockwaves from Chinese AI startups launching Kimi K3. Structural changes in the AI industry are already impacting broader capital markets—Zhipu’s stock fell over 38% in the two trading days following K3’s release, a transmission chain worth monitoring.
The knockout round for Chinese AI large models has just begun. The companies most likely to succeed in the next phase won’t necessarily be those with the largest parameters, but those that can convert model capabilities into real commercial value at the lowest cost and highest efficiency.
FAQ
Q: With Kimi K3’s parameter scale reaching 2.8 trillion, where does it stand among domestic AI large models?
Kimi K3 is currently the world’s largest open-source model by parameter count, and the first open-source model at the 3-trillion level globally. In comparison, Zhipu GLM-5.2 and Alibaba Qwen3.8 (2.4 trillion parameters) are smaller than K3. However, parameter scale does not equate to overall capability—K3 still trails closed-source models like Claude Fable 5 and GPT-5.6 Sol in comprehensive evaluations.
Q: Why did Zhipu’s stock price drop sharply after Kimi K3’s launch?
On July 17, the day Kimi K3 was released, Zhipu’s share price plunged 28.49%, wiping out over HKD 200 billion in market value. Market concerns stem from two factors: first, competition from open-source models may weaken Zhipu’s pricing power for closed-source models; second, Kimi K3’s performance in key scenarios like programming exceeded expectations, prompting a reassessment of Zhipu’s market share. As of July 20, Zhipu’s share price had fallen further to HKD 890.5.
Q: How does DeepSeek’s fundraising scale compare within the industry?
DeepSeek completed its first funding round of about $7 billion in the first half of 2026, with a post-money valuation of $52 billion. Just a month later, it reportedly launched a second round, with a pre-money valuation of $71 billion. This fundraising sets a record for single private placements by Chinese AI companies, with investors including the National AI Industry Fund, Tencent, CATL, JD.com, and others.
Q: What are the main challenges facing the commercialization of domestic AI large models?
The main challenges fall into three areas: compute bottlenecks—Kimi K3 paused new consumer subscriptions due to compute shortages after launch; paid conversion—the consumer side of large models is shifting from traffic competition to value validation, and the key issue is how to increase willingness to pay as user scale grows; profitability—the industry is still exploring sustainable business models, and transitioning from free consumer tools to enterprise-level paid services will require time for validation.
Q: Which is more competitive in the long run: open-source or closed-source models?
Both models have their strengths. Open-source models gain developer stickiness through low costs and ecosystem building, as seen with DeepSeek and Kimi K3. Closed-source models maintain profits via proprietary technology and API pricing, as with Zhipu’s GLM series. Current trends suggest more enterprises are adopting a mixed strategy: "open-source models for routine tasks + closed-source models for complex inference." In the long run, companies capable of operating both open-source ecosystems and closed-source commercial products may prove more resilient.
