Once again, the World Internet Conference Wuzhen Summit has arrived, and AI is undoubtedly the star of the show. Industry heavyweights have gathered on site, with both high-level dialogues and immersive experiences making the event truly packed with highlights.

This year's conference takes "open source and openness, co-building and sharing" as its core theme, and the AI open-source ecosystem has become the central thread of discussion throughout the event. From the dedicated forum "Artificial Intelligence Wuzhen Dialogue" to the announcement of 20 annual award-winning projects, a single list and a concentrated discussion together raise an industry proposition: as the AI industry shifts from single-point model competition to ecosystem co-building, the era when companies fought alone with closed-source models may be slowly turning the page.

Alibaba Chairman Joseph Tsai, for instance, spoke at the opening ceremony, saying that AI can only become widespread if it is open source, and only widespread adoption can create a sustainable and inclusive intelligent era. China's open-source models maintain a leading position in key indicators such as downloads and the number of derivative models, and the country is currently the global leader in open-source intelligence.

At the same time, among the list of leading scientific and technological achievements released this time, many projects point to cutting-edge hard-tech tracks. The "Jiuzhang No. 4" photonic quantum computing prototype made the list, further consolidating domestic technological accumulation in the field of quantum computing; Xiaomi's humanoid robot was also selected into the achievements list, and officials disclosed that it has achieved 8 hours of continuous operation and implementation in long-task household scenarios, marking that domestic embodied intelligence is no longer limited to stage demonstrations and has begun landing verification in real household scenarios.

The concentrated appearance of these two major directions—quantum computing and embodied intelligence—also shows from the side that the boundaries of AI competition are widening, and it is no longer just about comparing the conversational capabilities of large models themselves. Two or three years ago, the main theme of the domestic large-model track was closed-source racing. Major companies rushed to launch self-developed closed-source models, using parameter scale and evaluation benchmark scores as core promotional points. Model weights and core code were strictly locked, and capabilities and interfaces were all controlled by the vendors.

Users could only call APIs, and industry developers had no permission to modify or perform secondary tuning. The advantage of this model lay in strong vendor controllability and a clear commercialization path, but its shortcomings were equally prominent: high costs for small and medium-sized developers, severe homogeneous development, and industry innovation constrained by the iteration rhythm of a few leading companies. This year's Wuzhen Summit, however, put open source at the highest priority, clearly releasing a signal of industrial stage transition.

As AI develops to the present, the marginal returns brought by simply stacking parameters are becoming lower and lower. To penetrate deeper scenarios such as industry implementation, edge intelligence, robotics, and scientific computing, it is very difficult for any single company to complete this behind closed doors. The open-source model can open weights, code, and technical documentation, allowing universities, startup teams, and industry vendors to participate together in iteration, share R&D costs, and quickly fill scenario gaps.

From my observation, the open-source strategies of major domestic companies have already taken different routes, rather than simply opening everything at once. Some vendors choose to open-source the foundational base while keeping high-level capabilities closed source, using the open-source base to attract developers and monetizing through value-added services and industry private deployment solutions; some focus on open-sourcing lightweight edge models, targeting local deployment scenarios such as mobile phones, vehicles, and robots; and some research institutions prioritize opening papers and evaluation datasets to promote the establishment of unified industry evaluation standards.

Different degrees of openness directly reflect everyone's different explorations of open-source commercialization: embracing open source does not mean completely free; rather, it means switching to a different ecosystem game. Of course, under the open-source wave, contradictions and hidden dangers cannot be avoided, and this was also the focus of debate among many practitioners at the forum.

The first issue is intellectual property risk. There are many types of open-source licenses, and once developers mix open-source components under different licenses, authorization conflicts can easily arise, and subsequent commercialization and product launches may encounter legal disputes. The second is the challenge of security governance. After weights are open-sourced, models face the risk of malicious tuning and use for generating violating content. Open source does not mean letting things go unchecked, and the cost of model safety alignment will rise significantly.

The more realistic contradiction lies in commercialization. Open source can easily attract traffic and developers, but how to transform the developer ecosystem into stable revenue still has no standard answer. Many open-source projects are very popular but have long struggled to profit. Once major companies adjust their open-source strategies and tighten open permissions, the entire small and medium-sized development teams attached to them will be affected. The prosperity of the open-source ecosystem cannot be separated from long-term stable investment, and short-term traffic can hardly support sustained R&D investment.

A clear trend is that the future AI industry will neither move toward complete closure nor likely become an extreme situation of total open source. Most likely, open source and closed source will coexist for the long term and develop complementarily. General foundation models and lightweight edge models are suitable for open-source co-building, lowering the innovation threshold for the entire industry; exclusive models and high-end customization capabilities for scenarios with high security requirements such as finance and government-enterprise applications will still retain the closed-source model to ensure controllability and data security.

This concentrated discussion on open source in Wuzhen is not a simple update of industry slogans. It means the AI competition arena has changed: in the past, it was about whose model was stronger; in the future, it will be about who can build a more prosperous and sustainable developer ecosystem. That is, the next focus will mainly be on three chains:

Computing power chain: whether it can continuously and stably train and infer trillion-level models without relying on overseas high-end chips; ecosystem chain: whether open-source models, developer communities, tool chains, API prices, and the application ecosystem can form a positive cycle; commercial chain: whether AI can truly be embedded into enterprise operations, terminal interaction, and urban services, rather than remaining in demonstration scenarios.

In my view, the dividend period of closed large models is nearing its end, but the commercialization path of open-source AI has only just begun. Under the wave, whoever can balance openness and sharing, security and compliance, and commercial returns will be able to gain a firm foothold in the next round of the AI industry cycle. This is also what the capital market and the industry side truly care about. What do you think?