November 4, Wuzhen. The 2026 World Internet Conference Wuzhen Summit was under way, under the theme "Open Source, Open Access, Shared Benefits — Jointly Building a Community with a Shared Future in Cyberspace." In the halls, open source was one of the most frequently spoken phrases.
Zhang Yaqin was there — a foreign member of the Chinese Academy of Engineering and founding dean of Tsinghua University's Institute for AI Industry Research. In the line he gave reporters there were no parameters and no leaderboards, only the word "foundation." Xinhua later quoted it verbatim.
One Sentence in the Middle of a Report
"Open source and openness are, in themselves, the foundation of the internet," he said. "Openness produces vast amounts of data, which becomes the fuel for large models. As we enter the era of agents, the value of open source and openness becomes even more critical." The passage is short, but it sets the tone for a whole summit.
The sentence breaks into three layers. The first is a fact: the internet has been open since its birth. The second is cause and effect: openness brings data, and data feeds models. The third is a judgment: in the agent era, openness matters more.
The middle layer is the one most often skipped. Data does not fall from the sky; people upload it, bit by bit. Forums, encyclopedias, code repositories, question-and-answer communities — all of them grew out of openness.
What Open Source Actually Opens
For most people, open source means free. That is only half right. Open source means the source code is published, so others can read it, change it and redistribute it; whether it is free depends on the licence.
Open-source large models are a little different. What gets released is often not the training code but the weights — the numbers that make a model able to talk. Whether to release the weights, and how far, is another matter.
So the words "open source" carry different thicknesses in this industry. Some open only a calling interface; some release weights but forbid commercial use; some hand over the training details as well. The same phrase, very different content.
What Was on the Floor
In the summit's open-source ecosystem pavilion, several concrete things were on display: the 021 Science Foundation Model from Zhijiang Lab, the Hangzhou developer center of the ModelScope community, and national pilot bases for AI applications in six fields.
"Pilot base" is an awkward term. It refers to a place where something from the lab is put into real settings for trial production and trial use. Between a model that runs and a model that is genuinely useful, this is often the missing stretch.
In the same expo, Tencent's booth had Hunyuan Hy4 preview at its center. Staff said Hy4's weights would be opened in the near future, aimed at software engineering, research and intelligent office work. For a model, promising to release the weights at launch is itself a statement.
A booth is a good place to see what is real. Whether a model is good, one demo shows; whether a community is alive, the commit log shows. Compared with adjectives on stage, these objects say more about how far open source has come.
On November 2, the same summit handed its Leading Science and Technology Awards to 20 achievements, among them the "Jiuzhang 4" photonic quantum computing prototype and a vehicle-level agent platform for physical AI. The award list and the booths were saying the same thing.
Xinhua reported that the expo drew 671 companies from 54 countries and regions, with more than 1,000 new technologies and products. Open source was the undertone and the signboard. Floor space is limited; what gets shown has been chosen, and the criteria are themselves a statement.
The Track Moved Outward
The main forum added a session this year called the "Wuzhen AI Dialogue," gathering heads of global technology companies and open-source model firms along with AI scientists to discuss how to build a global open-source ecosystem.
According to Xinhua, a number of participants held that as foundation models become easier to obtain, the competitive track is extending outward: computing infrastructure, developer tools, industry solutions and service ecosystems are becoming the new battlegrounds.
That sentence deserves a second look. It means the model itself is becoming something like water and electricity — anyone can connect a pipe. What separates companies is how the pipes are laid, how the meter is installed, and who repairs them when they break.
This is also where open source is most often misread. Openness does not mean there are no barriers; the barriers simply move from "do you have a model" to "can you use it well," from a single point to a system. The first can be proven with one launch; the second takes years to show.
Open source is not charity for companies either. Release a model, and the more people use it, the more business there is in tooling and cloud services; hold the standard, and every later link is easier to negotiate. This is commerce, not sentiment.
The Weight of "Foundation"
Back to the word. A foundation is buried and invisible, but it decides how many floors can go on top. Over these decades of the internet, the protocols were public, the browsers competed, and the websites grew on their own.
The advantage of that arrangement is speed. Publish a standard and the whole world starts implementing it at once, and errors are found at once too. The disadvantage is mess: anyone can change a version, and convergence happens through de facto practice.
Large models are in something like the internet's position then. There are many models, each with its own interface, and for an agent to call a tool it has to be adapted one vendor at a time. The work is fragmented and heavy, and there is no common answer yet.
Zhang's judgment is that in the agent era, open source matters more. One way to read that: how strong a single model is, users can barely feel; whether tools interoperate and whether agents can work across platforms, they feel immediately.
Open source has its problems, of course. Once weights are out, they cannot be recalled; if someone uses them badly, the original author has no control. It is an old problem for every open-source project, only a model is harder to stop than a piece of code.
So open source and governance tend to be two sides of one thing. The more thorough the opening, the more someone has to set rules, run evaluations and keep records. That work produces no demo, but it decides whether the opening can last.
After the Halls Empty
The summit runs a few days; booths come down and forums disperse. What stays on paper is a few quoted lines and a few remembered names: 021, ModelScope, Hy4 preview. Names are remembered, lines are quoted, and the rest is up to time.
When the internet was young, public technical documents were not worth much; what was worth something was the people who turned documents into products. Today the order seems to have partly reversed: the documents themselves are becoming a kind of infrastructure.
In a hall where everyone talked about models, what he left behind came down to one word. It is neither new nor loud, but it decides how high the building can go. How high that is this year, no one can yet say.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.