The Open-Weights Gambit: How Lin Wei-cheng Is Betting China's AI Future on Giving Models Away

HANGZHOU — Lin Wei-cheng keeps a number taped to the wall behind his desk: the download count of his lab's flagship model, ticking past nine million as we spoke. It is, he says, the only metric he checks before coffee. Revenue, headcount, the closed rivals burning capital across town — all of that can wait. The downloads are the moat.

Lin, 38, runs Qingzhou AI, a lab he founded three years ago after a decade building search and recommendation systems at one of China's internet giants. Qingzhou does something that still strikes many investors as perverse: it trains models that rival the best closed systems in the country, and then releases their weights for anyone to download, modify and run for free.

The bet underneath that generosity is precise and, Lin insists, unsentimental. 'We are not an open-source charity,' he said, in a meeting room overlooking the Qiantang River. 'We are running the only strategy that wins in this market. Giving the model away is how we make the money.'

The logic of giving it away

Lin's reasoning starts from a reading of the Chinese market that he believes the closed labs have gotten wrong. In a crowded field where a dozen well-funded teams are within months of one another on capability, he argues, no single model can stay ahead long enough to charge a durable premium. The capability lead is real but perishable; the distribution is forever.

So Qingzhou races to be the model that developers reach for by default. Every download deepens an ecosystem — the fine-tunes, the tools, the documentation, the engineers who know its quirks — that a closed competitor cannot easily dislodge. By the time a rival ships something marginally better, switching costs have quietly accumulated on Qingzhou's side.

A closed model is a product. An open model is a standard. Products get beaten every quarter. Standards last a decade, Lin said, leaning forward as if the distinction settled the argument.

The money, in this telling, lives one layer up: enterprise support contracts, a hosted inference service for companies that do not want to run the weights themselves, custom fine-tuning, and a safety-and-compliance toolkit that Chinese firms need to deploy any model under regulatory scrutiny. The weights are free; running them well, at scale, under the rules, is not.

From recommender systems to frontier models

Lin's path to this conviction ran through a less glamorous discipline. At his previous employer he spent years on recommendation engines — systems judged not by benchmark scores but by whether they kept hundreds of millions of users engaged. That work, he says, taught him to distrust capability as an end in itself.

'In recommendations, the best model in the lab is worthless if it does not ship to every surface and improve every day,' he said. 'Distribution and iteration beat raw quality. I watched it happen a hundred times.' When he started Qingzhou, he simply applied the lesson to a new kind of model — and concluded that open weights were the fastest path to ubiquity.

The early days were lean. Qingzhou raised a modest first round when investors still assumed the winners would be closed, capital-intensive labs. Lin trained smaller, released aggressively, and let the download counter make his case. By the time the larger rounds came, the ecosystem argument was no longer theoretical — it was on his wall, ticking upward.

The compute problem he cannot wish away

For all his conviction, Lin operates under a constraint no strategy can dissolve: access to advanced chips. Export controls have made the most capable foreign accelerators scarce and expensive in China, and Qingzhou trains on a mix of constrained foreign silicon bought before the rules tightened and a growing share of domestic accelerators.

He is unusually candid about the trade-offs. The domestic chips are less performant and the software stack around them is younger, which means more engineering to get the same training run done. 'We spend effort on plumbing that our American counterparts get for free,' he said. 'But every quarter the plumbing gets better, and every quarter we depend less on chips we cannot reliably buy.'

He frames open weights, in part, as a hedge against that very constraint. A model the whole ecosystem improves is one Qingzhou does not have to advance alone on scarce compute. The community's fine-tunes and optimisations, contributed for free, stretch every GPU-hour the lab can secure. Openness, in his telling, is not just a go-to-market — it is a compute multiplier.

The closed rivals' rebuttal

Not everyone in Hangzhou's AI scene is persuaded. A founder at a rival lab that keeps its weights closed argued, when I relayed Lin's thesis, that openness is a strategy of the second-place. 'You give it away because you cannot charge for it,' he said. 'The frontier is held by whoever can afford the biggest training run, and that will always be a closed, capital-rich lab.'

Lin has heard the critique often enough to have a practised answer. The frontier, he says, is a moving target that no one holds for long, and a six-week capability lead is not a business. He points to the foreign open-weights releases that have repeatedly closed the gap with closed systems, dragging the whole field's prices down with them. 'They are doing my argument for me, globally,' he said.

The truth is that both men may be right about different time horizons. Closed labs can capture the premium at the bleeding edge; open labs can capture the vast market that forms just behind it. Lin is betting that the second market is the larger and more durable one — and that in China, where deployment runs through compliance and integration, the value was never really in the weights anyway.

What he is building toward

Asked where Qingzhou is in five years, Lin does not describe a bigger model. He describes a default. 'I want it to be unremarkable that a Chinese developer reaches for our weights, the way it is unremarkable to reach for a common database,' he said. 'When the choice is invisible, you have won, and you do not have to win again every quarter.'

It is a vision that depends on staying close enough to the frontier that the default never feels like a compromise — a tightrope between giving enough away to win distribution and keeping enough back to fund the next training run. So far the download counter on his wall keeps climbing, which is the only proof he says he trusts.

Whether the strategy survives contact with a market where capital, compute and policy can all shift overnight is the open question. But in a Chinese AI scene full of labs guarding their weights like trade secrets, Lin Wei-cheng has made a clarifying wager: that the surest way to own the future is to give the present away.