Summary of Key Points
Chinese open-source AI models have recently seen a significant performance surge. For instance, the model Kimi K3 from Yuezhi Anmian (Moon’s Dark Side) not only has its weights made publicly available but also surpassed American models like GPT-5.6 in programming benchmarks. Zhispu’s GLM-5.2 was rated by the US National Institute of Standards and Technology (NIST) as the “strongest open-source model” and is widely used by American companies. The gap between China and the US in AI technology has narrowed to just a few months, with Chinese open-source models becoming the de facto global standard. At this point, the US government is considering banning Chinese models, but the tech industry (including giants like NVIDIA and Microsoft, as well as startups like Replit) has issued a joint open letter opposing this move, as they are deeply dependent on these models for cost, security, and learning purposes. The direction of US regulation may shift from banning based on origin to controlling based on model capabilities.
I. The Rise of Chinese Open-Source AI Models: Outperforming GPT and Becoming Globally Popular
In the past, when it came to top AI models, people would think of American models like GPT and Claude. However, things have changed now:
- Outstanding Performance: After the release of Kimi K3’s weights, it immediately surpassed GPT-5.6 and Fable 5 in mainstream programming rankings, becoming the first Chinese model to do so. Zhispu’s GLM-5.2 was praised by NIST as the “strongest open-source model to date,” and even the US Vercel platform listed it as the fastest-to-adopt model for 2026.
- Widespread Adoption: Chinese open-source models, which can be downloaded, modified, and run independently, have become a fundamental part of the global AI ecosystem. Not only are Chinese companies using them, but American companies are also eager to adopt them. The gap in AI technology between China and the US has narrowed from years to just months, with Chinese models no longer being seen as followers but as competitors.
II. The Calculations of the Giants: Opposing Bans Is About Protecting Their Own Interests
On July 24th, 25 giants (later expanded to 133 companies, including OpenAI) jointly wrote a letter to the White House, arguing that banning Chinese open-source models would ruin the US AI ecosystem. Their reasons are practical:
1. Lowering Barriers to Entry: Open-source models allow small and medium-sized companies to develop their own AI applications without spending exorbitant fees on closed-source models like GPT. For example, a small company could create an intelligent customer service system by downloading and modifying Kimi without paying high API fees to OpenAI.
2. Preventing Monopoly: If only a few closed-source models (like OpenAI) dominate the market, the benefits of AI would be concentrated in the hands of a few large companies. Open-source models promote competition across the entire AI industry chain, benefiting more enterprises.
3. Greater Security: Closed-source models are like black boxes; it’s unknown if they contain vulnerabilities. Open-source models allow global developers to review them together, enhancing security.
It’s also important to note that they oppose a complete ban on “distillation technology” (the process of extracting useful information from one model to train another). While some in the US have accused Chinese models of “stealing” from closed-source models, the giants argue that legal distillation is an innovation and should not be banned outright. The implication is that while restrictions can be placed on Chinese models, the open-source model approach (like Hugging Face’s) should not be prohibited, as it would also affect the US open-source ecosystem.
III. Startups Cry for Help: Banning Chinese Models Would Kill Them
On July 22nd, 179 American startups wrote a letter even more explicitly, stating that banning Chinese models would lead to their demise. Nearly half of their products rely on Chinese models at their core (such as DeepSeek, Kimi, and Qwen). For example, the CEO of Replit said, “Banning Chinese models would be equivalent to banning open-source models themselves, setting a precedent for destroying the entire ecosystem.” The founder of Particle added, “If access is cut off, it’s not just about changing suppliers; our products’ foundational layers will stop working—our code is built on these models.” They also pointed out that using Chinese models saves millions of dollars compared to closed-source models. One company reported saving several million dollars by switching from Claude to DeepSeek.
IV. Three Reasons Why American Companies Rely on Chinese Models: Cost Savings, Survival, and Learning Opportunities
American companies rely on Chinese models out of necessity, not out of support for China:
1. Cost Savings: Companies are increasingly focused on costs. When a task does not require the most advanced model, they choose the best and cheapest option, which Chinese models often meet. For example, 80% of startups funded by a16z use Chinese models, and through platforms like OpenRouter, the proportion of companies using Chinese models exceeds 30% per week, with some using them up to 46%.
2. Emergency Solutions: Chinese models can be crucial when American models fail. For instance, when Hugging Face’s servers were attacked by an out-of-control OpenAI proxy, open-source models helped resolve the issue.
3. Learning Opportunities: American companies learn from the innovations in Chinese models. For example, the CTO of Arcee AI said they have gained valuable insights from Chinese models, and the new model Inkling from Thinking Machines Lab was trained using Kimi 2.5. This shows that even leading US labs are using Chinese models as a source of innovation.
V. Changing US Regulation: From Chips to Model Capabilities
The US government previously relied on banning chips (such as NVIDIA’s high-end GPUs) to restrict Chinese AI, but now that Chinese models have become a significant force, it needs to change its approach:
- New Regulatory Direction: The US Department of Commerce plans to issue temporary rules by September 30th (with public feedback). The focus may shift from banning based on origin to regulating model capabilities and access rights. Instead of banning entire countries (like China), the government aims to control model capabilities.
- Reasons for the Change: The industry opposes bans based on origin, so the government is seeking a compromise that limits the impact of Chinese models without damaging its own open-source ecosystem.
In summary, the rise of Chinese open-source AI models has made them indispensable in the global landscape. While the US government wants to ban them, the tech industry is opposed to this move. Ultimately, regulation may become more targeted, but the global prominence of Chinese models cannot be easily shaken by simple bans.