Summary of Key Points
July 2026 marked a critical turning point in the global large-scale model industry competition: leading companies both domestically and internationally (such as Zhipu, MiniMax, Yuezhiànmiàn, DeepSeek, JiéjuèXīngchen from China, and OpenAI, Anthropic, xAI from abroad) were rapidly iterating on their models, seeking funding, and striving to go public. The period during which a model could gain a performance advantage has shrunk from several months to just weeks, creating a cycle of "performance → market value/valuation → funding → even higher performance." In this "unlimited war," there are no permanent winners; each company must continuously invest money and iterate quickly to avoid being eliminated.
1. Model Updates Faster than Mobile Phones, with Leadings Lasting Only Weeks
In the past, large-scale model updates might occur annually, but now they happen weekly or even daily. For example:
- Just after Yuezhiànmiàn released its K3 model with 2.8 trillion parameters (the largest open-source model in the world), Alibaba's Qianwen announced a preview version of Qwen3.8 with 2.4 trillion parameters, DeepSeek V4 is about to be rolled out in a beta phase, and MiniMax' M3 Pro with 2.7 trillion parameters is also on the way. Elon Musk stated that xAI will release a new model every month, with Grok4.6 (2 trillion parameters) available by the end of the month, and he even claimed it would surpass K3.
- The day after K3's release, Zhipu's stock price dropped by 28%, and MiniMax' by 16%—this was not because of poor product quality but because the market suddenly realized that what was once considered the "number one domestic model" could be surpassed overnight. Once this assumption of leadership is shattered, valuations plummet.
In short, today's large-scale models are like new mobile phones: they become obsolete as soon as you buy them.
2. Going Public Is Not the End Point, but a Means to Raise More Money for Model Training
For large-scale model companies, going public is not an achievement; it's the beginning of a new phase:
- Zhipu raised HK$4.896 billion in its IPO and spent nearly 94% of that within a year on research and development. Therefore, in July, it issued another 314 million Hong Kong dollars in shares, using just 4.2% of its total capital to raise this amount, based on the market's belief that its company is worth nearly HK$1 trillion.
- After going public, MiniMax's market value dropped from HK$41 billion to HK$93.3 billion due to lock-up restrictions being lifted, but it immediately sought another HK$160 million in funding (through share placements and convertible bonds). The founder also announced a "zero salary policy until AGI is achieved," signaling to investors that he is fully committed and needs their support.
- Yuezhiànmiàn said it had no rush to go public six months ago, but now, with the release of K3, it is preparing for an IPO in Hong Kong because the valuations of Zhipu and MiniMax have soared to 500 times their sales revenue. This "window of opportunity" cannot be missed.
The core logic is simple: the higher the market value, the fewer shares are diluted during fundraising, leaving more money available for the next generation of models. Thus, all companies are striving to maintain high valuations.
3. Each Company Uses Different Strategies to Build Their Competitive Edge
Companies are exploring various ways to prevent being easily replaced:
1. Zhipu: Focusing on technological leadership and computing power, it announced that it will not pursue monetization for two years and acquired the domestic computing company Zhongke Jiahe to secure its technology and infrastructure.
2. MiniMax: Targeting overseas users and developing AI-native products; 70% of its 236 million users are from abroad, and its revenue comes from this market, allowing it to avoid domestic competition and seek growth in new markets.
3. Yuezhiànmiàn: Building an open-source ecosystem with the K3 model, encouraging developers to use it, thus creating a loyalty to their platform.
4. DeepSeek: Optimizing efficiency and costs; its V4 model has an API gross margin of over 50% (while others may still be losing money), and its prices are one-thirtieth of GPT's—using lower costs to attract corporate users.
5. JiéjuèXīngchen: Testing the feasibility of integrating hardware and software with the release of the AI smartphone STEPX Neo (without a price or release date), demonstrating its capability in more areas beyond just model development.
The common theme is that all companies are looking for additional capabilities beyond simply selling model APIs, as users focus on cost-effectiveness and will choose the best option available.
4. Price and Subscription Wars: Who Can Be More Aggressive?
If performance is not an advantage, companies compete on price and services:
- Domestic Price Wars: Major players have reduced prices six times this year, with three of those reductions being permanent. DeepSeek V4 costs HK$6 per million tokens (compared to GPT's HK$180), and MiniMax M3 was offered at half the price at launch.
- Overseas Subscription Wars: Anthropic has adjusted its subscription terms four times in two weeks, from a limited-time offer to a permanent subscription with a $100 bonus, while OpenAI removed the 5-hour usage limit on Codex and reset user quotas. xAI claims that Grok4.5 requires only one-fourth of the tokens used by Claude for tasks—all these measures are aimed at retaining developers.
The reason behind this is that computing power is a limiting factor, but losing users is even more detrimental. Every price cut or bonus offer is a way to attract new users, ultimately relying on fundraising to cover costs.
5. The Challenges Faced by Independent Companies Compared to Giants
Independent companies face greater pressure than giants:
- Giants' Advantages: Companies like ByteDance, Alibaba, and Tencent use their main businesses (e-commerce, social media) to fund model development without the need for frequent fundraising to justify losses. They are also shareholders of independent companies (for example, Alibaba has invested in Yuezhiànmiàn and MiniMax). If an independent company's technology is not leading, giants may withdraw their support.
- Opportunities for Independent Companies: They can focus more intensively and move faster. For instance, Yuezhiànmiàn's K3 model has more parameters than Alibaba's Qianwen, DeepSeek's inference efficiency exceeds that of giants, and Zhipu can temporarily match the performance of leading overseas models in certain use cases—these are things giants cannot achieve in the short term.
However, independent companies must move even faster: they need to raise funds as quickly as their competitors can iterate, or they will fall behind.
Conclusion: This War Has No End
July 2026 will be remembered not for a single winner but for a reality that everyone has accepted:
- No model will stop companies from seeking funding.
- No round of financing will prevent them from continuing to innovate.
- No price will deter competitors from pursuing advancement.
It's like an endless marathon: once you cross the finish line, the next race begins. Everyone must be prepared at all times for the next challenge. This is the nature of the "unlimited war" in the large-scale model industry.