The 2026 World Internet Conference opened in Wuzhen on November 3 with a theme that would have been hard to imagine a few years ago: open development and shared benefits, working together to build a community with a shared future in cyberspace. Open source, long treated as a side topic in a breakout session, was pushed to center stage this year. Two days before the opening, China's first open-source-themed ecosystem exhibition hall opened its doors; the main forum, a dialogue on artificial intelligence, put the global open source ecosystem at the core of its agenda; and the Wuzhen startup competition added a dedicated open source track for the first time.
There is also a tangible layer to the venue itself. Walking through the exhibition hall, visitors encounter open source not as abstract repositories but as running systems: a scientific foundation model on a screen, a pharmacy robot moving down an aisle, database deployments with counters still ticking. Open source was rendered physical — something you can watch work, not just clone. And behind the stands, an old pattern is being rewritten: the same teams that spent a decade using other people's open source are now the ones publishing their own foundations.
Putting open source at center stage was not a gesture of sentiment. It rested on a report card that can be checked item by item: officially reported figures, third-party platform data, and cross-confirmation from industry reports. Laid out together, these numbers sketch where Chinese open models actually stand — what has been accumulated over years, what is being consumed right now, and which new frontiers have just been opened.
A Report Card: Where the Numbers Come From
The heaviest set of numbers came from the opening ceremony. Alibaba co-founder and chairman Tsai Chongxin said at the opening ceremony that Qwen, Alibaba's model family, has released more than 470 open-source models, with cumulative global downloads exceeding 4 billion and more than 330,000 derivative models built on top of them — all three figures ranked first worldwide. He added his own diagnosis: Chinese open models combine strong overall performance with high inference efficiency, and they continue to lead on key indicators such as download volume and the number of derivative models.
The framing matters as much as the figures. Tsai's argument was that AI can only become universal through openness: the benefits of intelligence should not sit behind a single company's paywall. The three first-place rankings were offered as evidence that this approach compounds — the more people build on an open base, the more useful, the more tested, and the more efficient that base becomes for everyone who follows.
The most revealing figure in that list is the 330,000 derivative models. Downloads can come from curiosity; derivatives can only come from reuse. A base model earns derivatives only when it is good enough, its license open enough, and its documentation complete enough for others to modify it into something of their own. A count in the hundreds of thousands means Qwen is no longer just Alibaba's product but a shared foundation that developers around the world keep rebuilding.
Third-party data tells the other half of the story. OpenRouter data shows that in the week of October 19 to 25, DeepSeek V4.1-Flash-1017 handled 33.6 trillion tokens of traffic in a single week, ranking first globally. During that same week, the top five models by global call volume were all Chinese, and their combined traffic overtook that of American models. OpenRouter is one of the main channels through which developers worldwide aggregate their model calls, so this is not a vendor's self-description; it is a vote cast with real invoices.
Industry reports cross-confirm the picture. According to the 2026 Generative AI Application Development Report released by the policy and international cooperation institute of the China Internet Network Information Center, domestic companies such as DeepSeek and Moonshot AI have released multiple trillion-parameter open-source models, and the top six models on global mainstream call leaderboards all come from Chinese teams. Zoom out one more level: according to figures cited by Zhejiang's official publicity platform, cumulative global downloads of China's open-source AI models have surpassed 10 billion. Official self-reporting, third-party statistics, and industry reports now point to the same conclusion.
Center Stage Is Not Free: Where Wuzhen Put Open Source This Year
Moving open source from a breakout-session topic to center stage took three concrete arrangements, not just a slogan. The first is the exhibition hall itself: jointly built under the lead of the Cyberspace Administration of China and Xinhua, the country's first open-source-themed ecosystem hall opened on November 1, bringing open source out of code repositories and community forums to where ordinary visitors can see what the ecosystem looks like. The second is agenda design: the main forum focused on the global open source ecosystem, and the startup competition added an open source track so that teams can compete and be judged on their open work. The third is an official statement: a leader of the cyberspace administration said at the conference that interoperability is the natural gene of the internet, that open source and openness should be encouraged, and that innovation should benefit the whole world.
That statement deserves a separate read. Open source in China used to be a spontaneous behavior of developer communities; written into the top-level agenda, it now comes with aligned resources, policy expectations, and industrial signals. Developer communities are most sensitive to this kind of shift in the wind: when the conference theme, the exhibition halls, the competition tracks, and the regulator's statements all point the same way, open source no longer has to prove its value to every company one project at a time.
An even stronger signal is a change in who is doing the opening. The old pattern was "using other people's open source." The new pattern is opening up one's own foundations. Zhejiang Lab's 021 scientific foundation model, the ModelScope community, Ant's Bailian platform, and OceanBase — open source for five years — share one trait: they tear down thresholds. According to a commentary from a Zhejiang television program, OceanBase has released more than 5 million lines of core code over five years of open source, with more than 200,000 deployments. A database company willing to publish its core code was almost unthinkable in China's software industry a decade ago.
Call Leaderboards and Download Leaderboards: Two Kinds of Truth
To read China's open source score honestly, two leaderboards should be kept apart: the download leaderboard and the call leaderboard. They tell two different truths, and blending them leads to overestimating on one side and underestimating on the other.
Downloads are accumulated momentum. More than 4 billion downloads and 330,000 derivative models show that the Qwen family has been adopted worldwide as a base to build on. This resembles the balance sheet of an ecosystem: it records years of accumulated trust — who is using it, how deeply, and how much of their own work they are willing to put on top. The strength of accumulated figures is their stability; the weakness is that they cannot answer how many people are actually using the models right now.
Calls are present-tense usage. The 33.6 trillion tokens that DeepSeek V4.1-Flash-1017 processed in a single week, and the fact that the top six models on mainstream call leaderboards all come from Chinese teams, mean developers around the world are consuming Chinese models' inference capacity with real money at this moment. A download can be one-time curiosity; a call is a recurring bill. Nobody pays long-term for a model that does not work. Traffic climbing to the global top means the models are not just downloaded but wired into real business processes.
There is also a practical lesson in how the two leaderboards move at different speeds. Download counts grow slowly and fall slowly; they reflect decisions made months or years ago. Call volumes can swing within weeks as new versions ship and workloads shift. That is why the two need to be read together: the slow number shows whether an ecosystem has roots, while the fast number shows whether those roots are still drawing water.
Read together, the two leaderboards complete each other: accumulated momentum and present-day usage are both online, which is what turns the story from a single viral release into an ecosystem with continuity. A single hit can be displaced by the next release; a continuous ecosystem locks developers in through switching costs. That is the part of the Wuzhen report card that stands on firm ground.
Scientific Models and Embodied AI: New Open Source Frontiers
If language-model open source has entered an ecosystem competition, this year's Wuzhen displayed two newer front lines: one going deeper into science, the other into the physical world.
The first is scientific models. Zhejiang Lab's 021 scientific foundation model was displayed in a prominent spot in the exhibition hall. The name itself carries the meaning of going from zero to one: it is the first domestic exploration of an AI foundation model built for scientific research. According to introductions at the site, Zhejiang Lab has completed the full training of a 236 billion parameter model and released it online. It handles scientific data across modalities such as images, sequences, and spectra, filling a domestic gap in AI foundation models for science. Scientific data is complex in modality, scarce in labels, and expensive to work with. Open-sourcing the result of such an expensive path means research teams across the country can start from that baseline instead of feeling their way from zero.
The second is embodied AI. LingBot, the robotics model series from Ant's Lingbo team, demonstrated a smart pharmacy scenario in the exhibition area: inside a narrow aisle 80 centimeters wide, a robot independently completed order intake, intelligent sorting, and product delivery. More interesting than the single demo is the approach — an embodied-native, cross-embodiment foundation model designed to let one model adapt to different robot bodies. According to data shown at the booth, the LingBot series has surpassed 2.2 million downloads in the nine months since it was open-sourced. Open-sourcing embodied models is far harder than open-sourcing text models: data must be collected in real scenes and validation must run on real hardware. That download figure shows real developer demand flowing into this new front line.
Tencent's booth placed its bet on productivity scenarios. The Hunyuan Hy4 preview model was fully open-sourced at release, targeting software engineering, scientific research, and intelligent office work, with a focus on understanding, planning, and verification for long-horizon tasks. Long-horizon tasks are precisely the bottleneck for agents today: a task decomposed into dozens of steps requires the model to understand, plan, and verify itself along the way. Releasing that capability as open source hands one of the hardest stretches of agent engineering to the community to pave together.
Scientific models open source to fill a gap; embodied models open source to test a direction; productivity models open source to compete on deployment. The common thread is that Chinese teams no longer open-source only one category — language models — but treat open source as the default release channel for every frontier.
What Ordinary Developers Actually Get
The report card and the leaderboards are an industry view. For an ordinary developer, the more practical question is: what does any of this do for my project?
The answer lives outside the exhibition halls. When the density of open models is high enough, a small project with no compute budget and no team can start from the same model foundation as the giants. Selection cost approaches zero: models are picked from open communities, documentation and fine-tuning recipes are ready-made, and the people who hit the same pitfalls are one forum thread away.
The economics changed on the tooling side as well. Open toolchains — inference frameworks, fine-tuning scripts, evaluation suites — mean a single developer can carry a project from prototype to deployment without buying enterprise licenses at every step. What used to be capital expenditure becomes community knowledge, available to anyone willing to read the documentation.
According to a report by Tonghuashun, a financial data platform, Chen Yu, CTO of Yuanzhen Digital Technology and a core author of the open-source project llm-self, put it plainly: without open models and an open toolchain, he would spend several times more time and effort, and the project would most likely fail anyway. He runs a one-person company — and precisely because the foundation is open, a company of one dares to take on work that once required an entire team.
Read in reverse, his words carry more weight: when 470 open models, 10 billion downloads, and 33.6 trillion tokens of weekly traffic form an open foundation, one person can be a whole team. Wuzhen's decision to put open source at center stage will ultimately be tested not by its agenda or its booths, but by whether developers like Chen Yu can get their ideas running on that foundation. If the answer is yes, this report card is more than numbers.
Related reading