Summary of Key Points
Recently, leading AI companies both domestically and internationally (such as Alibaba, DeepSeek, Meta, xAI, etc.) have been actively releasing or updating large-scale open-source models, with a focus on enhancing the capabilities of these “agents.” Open-source models have shifted from being followers of closed-source models to posing a competitive threat to them. For the first time, the number of downloads of Chinese-developed open-source models has surpassed that of American models. Open-source models offer advantages such as local deployment and lower costs, which squeeze the pricing power and profit margins of closed-source vendors. Closed-source companies are responding with strategies like “layered open-sourcing,” while the United States faces disadvantages in the open-source AI competition due to policy restrictions.
Detailed Analysis
1. Collective Upgrade of Open-Source Models: Emphasizing Autonomous Functionality
The core of these updates is to make AI more autonomous—like employees who can carry out tasks on their own. For example:
- The programming intelligence score of DeepSeek V4 Pro has soared from 12.8 to 62.7, indicating that it can write code and solve programming problems more independently;
- Alibaba’s Qianwen model can program autonomously for 16 consecutive days and can even replicate scientific research papers or manage a virtual e-commerce store (with over 2,000 interactions without errors);
- Meta’s Muse Glimmer has been optimized to handle the entire process from planning to error correction. For instance, it can find data, analyze information, and write reports on its own, and it will self-check and correct any mistakes along the way. In simple terms, AI models have evolved from being question-and-answer assistants to becoming capable of actively planning and executing complex tasks.
2. Reversal in the Status of Open-Source Models: China Leads in Downloads
Open-source models, which were once followers, are now challenging closed-source models. Data from Hugging Face (the world’s largest AI model platform) shows that Chinese-developed open-source models account for 41% of downloads, surpassing those from the United States for the first time. Market research firm Counterpoint suggests that if Western giants continue to rely on closed-source models (like creating walled gardens that prevent outsiders from accessing them), developers and companies will turn to Chinese open-source options, as they can be freely downloaded and modified without relying on others. This indicates that China has moved from following in the footsteps of open-source AI to competing on equal terms or even leading the field.
3. Open-Source Models Making it Difficult for Closed-Source Companies to Profit
The cost-effectiveness of open-source models poses a significant threat to closed-source vendors:
- Profound Cost Differences: For example, the output price of the open-source DeepSeek V4 Pro is less than one-seventh of that of the closed-source Grok 4.6. Closed-source models require payment per usage, while open-source models can be deployed locally and used for free or at a low cost, significantly reducing the cost of implementing AI solutions;
- Loss of Revenue and Ecosystem: With more developers using open-source models, closed-source companies not only lose revenue from API calls but also miss out on the “data feedback loop” (the more users there are, the more data is generated, and the better the model becomes). Without user feedback, the iteration speed of closed-source models slows down;
- Declining Valuation: Investors see limited future profit potential for closed-source companies, leading to lower long-term valuations.
4. Closed-Source Companies’ Strategies: Layered Open-Sourcing or Reintroducing Open-Sourcing
Closed-source companies are finding ways to adapt:
- Reintroducing Open-Sourcing: For instance, Meta has slowed down its open-sourcing efforts before but is now pushing it again (with the release of Muse Glimmer and hints at making its flagship models more accessible);
- Layered Open-Sourcing: xAI has made older versions (Grok-1, 2.5) open-source to attract developers and gather feedback, while keeping newer versions (Grok 4.6, 4.7) closed-source to maintain its core technical advantages. This approach allows them to build an ecosystem while protecting their key technologies.
5. Policy Barriers for Open-Source AI in the United States: Zuckerberg Calls for Relaxation
The United States faces inherent challenges in the open-source AI competition:
- Zuckerberg has criticized strict regulations on data use and model training, which limit the development of local AI labs and put them at a disadvantage compared to overseas teams (such as those in China);
- He believes that banning foreign open-source models is counterproductive and will prevent the U.S. from maintaining its leading position. He advocates for policy changes to simplify data usage and model development processes, allowing U.S. open-source AI to continue to lead. These policy restrictions have become a stumbling block to the development of open-source AI in the country.
Conclusion
Open-source AI is becoming the new standard in the AI industry, offering performance on par with closed-source models at lower costs and with greater flexibility. China has gained an advantage in this field, while the United States faces policy-induced challenges. In the future, the competition for open-source models will intensify, forcing closed-source companies to either transform or struggle to survive. For businesses and developers, the widespread adoption of open-source models means that the barriers to using powerful AI tools will continue to decrease, making advanced technologies more accessible to everyone.