China’s Open-Source AI Push Tests The Closed-Model Playbook
Former Hugging Face Asia-Pacific ecosystem lead Tiezhen Wang said Chinese AI labs are using open releases, licensing changes and cheaper token economics to challenge closed U.S. model strategies without relying only on direct model fees.
China’s AI Labs Lean Into Open Releases
Chinese AI labs are treating open-source model releases as a strategic route into the global developer ecosystem.
Tiezhen Wang, formerly head of the Asia-Pacific ecosystem at Hugging Face, framed the divide as less about a simple U.S.-China race and more about different engineering incentives around model access.
The core contrast is clear.
OpenAI and Anthropic keep proprietary model code behind commercial interfaces, while Chinese labs are releasing models that developers can inspect, customize and run without depending entirely on a foreign platform.
Wang said releases from Chinese labs are already useful to U.S. researchers, including DeepSeek reinforcement learning training work and Chinese open weights running on U.S. hardware.
The Monetization Question Is Not Settled
Open releases do not remove the business problem.
Wang said labs can struggle to earn money directly from a model once it is released, but they can still build revenue around infrastructure, subscriptions, application programming interfaces and retained base models.
Kimi was his example of that split.
The company released a model for free, yet demand remained strong for its interface and subscription because users still need reliable infrastructure.
Wang also said a lab can open a fine-tuned model while keeping base models private for commercial sale.
That model gives open-source labs a different kind of leverage.
The first advantage comes on day zero, when the research team behind a release already knows how to operate it.
The second advantage is reputation: strong public models help young labs recruit researchers and become visible to developers before they have a mature paid product.
Licenses Are Becoming The Control Layer
The open-source strategy is also becoming more selective.
Wang pointed to Minimax changing its license so commercial users that make money from a model have to pay.
His concern was not individual developers using a model for free, but cloud providers running it for profit without sharing value with the lab that created it.
That licensing shift matters because it is a middle path between fully closed models and unrestricted free use.
A lab can preserve developer access while asking revenue-generating infrastructure companies to contribute.
If Chinese labs cannot find that balance, Wang warned that monetization pressure could push more of them toward closed-source strategies.
Capital markets are another support line.
Wang cited Zhipu’s stock price as having reached 10 times growth, arguing that investor backing can help labs buy compute, recruit talent and secure data.
The point is not that open-source automatically pays for itself, but that financing can keep model labs active while they search for sustainable revenue.
Token Costs Shape Adoption
For startups, Wang described a pragmatic path rather than ideological loyalty to either open or closed models.
A company may begin with a closed model that fits the product, gather users and data, then later consider switching to an open-source model to cut token expense.
He said that move can eventually save maybe a hundred times on tokens.
The claim reflects the operating pressure behind the open-source debate: model access is not just a research preference, but a cost structure that affects whether products can scale.
Wang also contrasted U.S. and Chinese adoption patterns.
He said Uber burned through an entire year’s token allocation within four months, while Chinese internet companies are pushing employees to use large token volumes because domestic open models are cheaper to run.
What To Watch Next
What matters now is whether Chinese labs keep releasing useful models while tightening commercial licensing for cloud providers.
If that balance holds, developers get more model choice and labs get a path to revenue.
If it fails, the strongest labs may reserve more work behind paid interfaces.
For enterprise AI buyers, The practical question is narrower: whether open models can reduce token costs without creating new risks around support, governance, intellectual property and operational reliability.
Wang’s comments leave one clear measure to watch: whether cheaper Chinese open models cross the threshold where companies use them in everyday products, not only research experiments.
