How Do You Build an AI App That Outruns Every Model Release?

When a new foundation model ships, it can wipe out the value of an entire AI application overnight. The same model that makes an image-generation app possible is one release away from making it free. For independent application companies built on someone else's model, every quarterly model drop is an existential event.

Chen Mian has built his whole company around surviving that cadence.

Evoken, the parent of the AI image tool Liblib, the design agent Lovart, and the video app LibTV, pulled in more than $300 million in annual recurring revenue across its three products. That is a rare scale for a Chinese AI application studio that does not own a foundation model. During the worst of the AI application funding winter, the company closed a $300 million round at a $2 billion valuation, a bet that the team could keep rebuilding faster than the model layer kept moving.

The model release is the recurring forced-pivot trigger. Chen Mian describes the cycle on LatePost's podcast Episode 175: a new model lands with a capability the app was built around, the app's value compresses, the team has to find the next surface the model has not yet absorbed. Liblib started as a Stable Diffusion model community; Lovart became a vertical design agent when the underlying image model could already do what general tools did; LibTV rode a video generation gap that earlier coverage tracked as a near-term opportunity for fast followers.

The pivot cadence is the company.

Outside investors, ex-employees, and the Chinese tech press have all filed versions of the same accusation: Evoken is fast-follow rather than original, runs on aggressive paid-traffic spending and discount pricing rather than product pull, and is one bad quarter from a public "blowup" (爆雷). Some former staff told reporters the company is not technical enough, not innovative enough. Co-founders have left. PingWest's profile of the LibTV launch noted the discount-tier pricing and the originality questions that followed. A 36Kr English piece on Lovart captured the same tension: vertical AI agents in China are racing a model layer that does not slow down for them.

Chen Mian hears all of it. His counter-frame, laid out across the two-hour conversation, runs like this: this is not a self-destructive play, and the unit economics are not negative gross margin. The LibTV 3.9 discount membership tier, for instance, was built for a specific user need, not a generic subsidy. Innovation has a cost application companies cannot afford to ignore. His own line, "不侵略性强，我怎么厉害呢？我不厉害，我怎么活下来呢？" ("If I'm not aggressive, how can I be strong? And if I'm not strong, how do I survive?"), is meant as a working theory, not a posture.

The management frame he offers in return is a "two reward models" structure: explicit rewards for the work the company has already decided to do, and a parallel lane for the co-founder-grade free exploration that produced the next pivot. He admits in the same interview that the gap between first product-market fit and actually winning the market is large, that he has cried in front of shareholders before a launch (Lovart), and that the deeper bet is whether humans still have value in a loop where the model keeps absorbing the previous product surface.

The structural read for anyone who depends on an AI app: every product is hostage to a model release cycle the company does not control. Watch four things. A product with one core capability is the most exposed when a new model can do the same thing in the same prompt. A business whose margins track a single model's API price has its unit economics rewritten by one repricing. A team that cannot pivot without melting will not survive the next forced-pivot trigger. And a company stuck in the gap between "first to find product-market fit" and "wins the market" has to outrun model releases for more quarters than its revenue can compound.

Evoken is the case study, not the verdict. The $2 billion valuation is a bet that the team can keep beating the model release clock; the controversies are a reminder that the bet is open. Either the playbook generalizes, with speed, anxiety, and a willingness to throw out last quarter's product, or the model layer eats the application layer first.

Chen Mian's own closing image, after three years of running Evoken at full throttle: he has been driving a very fast car, and only now is starting to look at the road.