Liblib, what's your dream?

@吴睿睿

A few days ago, Liblib — oh wait, Evoken (Yanyu Technology), announced a new round of funding.

Over the past three years, this AI application company has cycled through at least eight names: Liblib, Lovart, LibTV, Xingliu, Zaoci, Shakker, Evoken... Before Evoken, there was even a transitional holding company name. The profusion of names reflects the profusion of products. And every time one of those products found success, the company's valuation and funding would climb another rung.

Evoken, officially, evokes creativity. But it sounds an awful lot like "Evoke Token" — the true token awakener.

Yet as one of the top three Chinese AI application companies by valuation (the other two being Manus and Genspark), Liblib may also be the most controversial.

Despite already being "China's largest AI creation platform," many people's first reaction upon hearing its name is still "token relay station": users come here to access the fastest, cheapest, most comprehensive models. This reputation has fueled industry lore that Liblib subsidizes its products to the point of negative gross margins — a figure that should normally be above 70%.

Even within the investment industry, whether you invested or passed is itself a statement of position. But in my perception, few companies provoke such intense bewilderment, even outright criticism, from those who passed as Liblib does. At one fund that did invest, another partner told us: "This company will blow up eventually."

That hasn't stopped the chase. This round's post-money valuation exceeds $2 billion — four times its Series B valuation just eight months prior.

These are merely the surface-level issues. A more accurate framing might be: an AI company that seems to be following the old internet playbook; a product company whose every product carries traces of imitation; a startup that burns cash competing with giants.

In the AI era, this is a company that isn't very AI, yet has achieved outsized success.

Existence implies rationality. We have no intention of scrutinizing a startup under a microscope, but we are genuinely curious how such a contradictory story came to be. elsewhere spoke with more than a dozen people close to Liblib, attempting to answer this question.

Given that Evoken hasn't yet achieved any real name recognition, we'll refer to it throughout by its original name: Liblib.

A Manufacturing Company

When you think of an AI company, what's the first thing that comes to mind?

Perhaps something like DeepSeek, which found a new path for large models, or Manus, which first defined what a general-purpose agent could be. These imaginings likely share a common thread: creation.

And this, precisely, may be what Liblib possesses least of all.

Looking at the company's three most important products today: Liblib is China's version of Civitai; Lovart was inspired by Manus; LibTV and TapNow are pixel-for-pixel similar.

In other words, aside from Lovart as the "world's first creation agent," this is a company with little originality.

When many technical hires join Liblib, their two most important tasks are: one, ensure API stability; two, integrate new models. Veteran employees tell them: "No matter what product features you build, none will bring in as many new users as integrating a new model. Just wait for the next model release."

The department most eager to integrate new models at Liblib isn't product or engineering — it's marketing, because they're rushing to distribute press releases about being "first to integrate XX model." The two technical teams, based in Beijing and Shanghai respectively, even compete for marketing's praise based on who performs better at model integration.

A former Liblib employee responsible for model integration told me that every time a new model dropped, Liblib's speed ranked in the top three — sometimes even faster than Dreamina when integrating ByteDance's models.

He also said that to burn less money, Liblib sought out numerous relay API channels. Marketing staff were responsible for sourcing: perhaps a major customer's bulk protocol package going unused and resold cheaply; perhaps from specialized suppliers like FAL.

The most frequently mentioned example was their bulk purchase of discounted Nano Banana accounts. To secure better prices, many of the accounts they procured were student accounts, because Google gives eligible students over a year of premium model credits.

In short, a thousand threads to gather, then the technical team executes one by one: this channel has a 25% discount, cheaper than our current 20% off, should we migrate some traffic over?

So some say Liblib's essence is a wholesale-to-retail business.

Hence the inevitable occupational hazard: counterfeit goods. Recently Liblib has faced considerable public sentiment around fake models, with users complaining that cheap substitutes were used in place of Nano Banana and Seedance 2.0. A former Liblib technical team member told me the actual issue likely stems from relay APIs — often they couldn't even trace where the source was.

There was one instance where a major supplier delivered fakes, causing Chen Mian to explode in anger and demand accountability within the company.

Every day, in every work group, Chen Mian drives this nearly 200-person company: competitors launched a new discount package, we need to match it fast; one approach isn't working, pivot 180 degrees immediately; there's a conversion funnel issue, commercialization and R&D are both pulled from bed to work overtime. Since founding, the company has operated on a six-day workweek.

If instructions go out and aren't advanced that day, that's too slow. Chen Mian's follow-ups appear in the group chat. He habitually sends short messages, each less than 2 seconds apart, popping up one by one to fill the screen: I said this last time — why hasn't it changed — just leaving it — letting users see this? Then @A, @B, @C.

Hard to imagine: the operating principle of a top 3 Chinese AI application company seems far from creation, yet remarkably close to moving bricks.

It resembles something more traditional. I once asked a Liblib executive what that might be. After considerable thought, three words: manufacturing.

Indeed, manufacturing's underlying code is: product innovation doesn't matter, what matters is cost control, worker management, and maintaining supply chain relationships.

An employee created a derivative of a Japanese salaryman meme, with a red-faced, enraged Chen Mian shouting: at this rate the company will go bankrupt! Many employees loved using it; Chen Mian himself would send it too.

Revolving Door CTOs

When you think of an AI company, what's your second assumption?

I suspect: technology?

But nearly everyone will tell you — even current Liblib employees will tell you — this is not a technical company.

A former Liblib technical employee told me that in just three years since founding, Liblib has gone through seven CTOs/technical leads. In his view, Lovart's slowing data growth since late last year was directly tied to the CTO's diminishing authority.

Interestingly, from an investor who backed Liblib, the narrative becomes: (that) CTO wasn't top-tier, and has been replaced.

Liblib's breakthrough owes most to Lovart: the company once possessed an AI-native team, and proved it could build innovative products.

Manus made Chen Mian realize that next-generation products needed to be built on algorithms, so for Lovart, Liblib established a more AI-focused technical team in Shanghai: some from Tencent, ByteDance, and Alibaba's AI labs; some from Doubao and MiniMax's product teams.

With different DNA and distance from headquarters, this team briefly developed a new collaboration model. The traditional "product writes PRD, brings to engineering" structure was dismantled. Ideas no longer flowed only from product; engineering would proactively approach product to find the right interaction forms for their technology. The entire team finalized the plan before dividing into functional execution.

A Shanghai-based executive at the time believed this was the most important reason Lovart achieved such good results in so little time.

But soon, the relationship between Beijing and Shanghai teams grew delicate.

Chen Mian and the Shanghai team held diametrically opposed views on "how to build an agent." To the Shanghai team, an agent is a person; what it needs isn't correcting its actions, but exploring how to unleash its intelligence. The director leading Shanghai's technical team, from Tencent's Youtu Lab, wanted to continue the integrated product-engineering model, encouraging engineering side projects for product innovation.

But what Chen Mian needed most now was stable user delivery, continued revenue growth for Liblib and Lovart. Whether it was an agent or a generator mattered little.

Those who joined Lovart in late 2025 came with aspirations of exploring image AGI. One frontend Twitter star broke his never-commute rule and endured three months of commuting; another rejected a 80,000 RMB monthly offer from Doubao, taking a 50% pay cut to join Lovart. The brick-moving work was dirty work to them; product writing PRDs and engineering executing was regressing from AGI.

Several times, Chen Mian's instructions couldn't be implemented by this team, which he found unacceptable.

In a 70-person group, in all-staff emails, Chen Mian attributed Lovart's later poor data performance to "the technical lead's vision not aligning with mine," forcing the other party to pledge to dock their own year-end bonus.

Multiple Liblib technical and HR sources told me that, to date, Lovart's original technical team has almost entirely departed.

The replacement came from Doubao, internally codenamed Stone, with many years at Taobao. Upon taking over, his instruction to the team was — abandon all side projects, follow orders.

"None of our original team understood technology," a Liblib executive told me, and in areas he doesn't understand, Chen Mian struggles to trust others, making CTOs "very difficult to land."

Some investors believe Chen Mian's iteration speed is remarkable, and with money as a moat, when the next phase demands stronger technology, he'll certainly recruit a stronger CTO. Some Liblib executives have expressed to us their need for someone like Peak Ji, with both technical and engineering capabilities.

Indeed, Peak to Manus is clearly irreplaceable. Only unknown: if Peak had joined Liblib, could he have successfully landed?

Sense of Rhythm

But on the other hand, if we look only at growth data, Liblib's report card is genuinely impressive.

Multiple investors who saw Liblib's fundraising data pack told us it was "hard not to be moved."

The numbers released in this round's publicity were striking: ARR exceeding $300 million; LibTV daily revenue surpassing $1 million; Liblib AI cumulative users over 30 million, Xingliu (Lovart's domestic version) users exceeding 10 million.

But notice: new products announce revenue data, old products talk user numbers.

A person close to Liblib's core team told us that with every new product launch, the company tilts all resources to pump up the numbers and crafts the narrative. Pretty data always excites investors; the only anxiety is, "what if investors ask why the previous product's numbers dropped."

Old products withering when the money and resource IV is removed is certainly problematic. But this seems to be a rhythm Liblib has gradually mastered: build new product, get new growth data, raise more money. At least for the past three years, Chen Mian has proven this playbook fits the current technology cycle. Before model convergence, all intermediate products are disposable supplies.

In some supporters' framing, this is like surfing: essentially about catching opportunities; painstakingly crafting products risks being swallowed by models.

More than one investor told elsewhere: LibTV's borrowing from TapNow is a "good signal." Since "world's first XX" products can't always come from Liblib, let smaller startups validate first, then overwhelm.

"Surpassing the original in a week or two — when he wants to enter other markets in the future, Chen Mian will also be the winner."

If money is key to improving the odds of the next surf, he should have more. An investor who calls himself "lucky" told us: "Occupying this multimodal stronghold, getting more money, more attention, recruiting better people, leaving nothing around you, creates hope of surviving to the end."

An internal source told us Chen Mian also considered acquiring TapNow.

But note: Manus and Liblib are actually quite different. Manus grows with cleverness, and exits quickly while ahead — guerrilla warfare against giants. Liblib buys growth with cash, fighting giants and small companies alike. In a sense, you could say: Liblib itself is a giant — just relative to smaller companies.

Late last year, Liblib began more aggressively recruiting ByteDance senior staff, offering salaries competitive with ByteDance. HR began building shared services. The company rapidly expanded from 40-plus to nearly 200 employees, adding more than a dozen monthly. Like ByteDance: hire massively, cull swiftly.

Awakened by Manus

Many call Liblib a club deal. This isn't quite accurate. Until late October last year, Chen Mian wasn't a star CEO in primary markets.

In 2023, Liblib's first round valued the company at just $18 million. For investors Gaorong Ventures, Source Code Capital, and GSR Ventures, there was little to quibble about at that price. It was Chen Mian who later complained repeatedly: he didn't understand fundraising then, letting three funds in simultaneously at a low valuation.

The second round was harder. Money was nearly exhausted by subsidy wars, and the product had been taken down. Baidu's strategic investment provided timely aid, but with conditions — only $5 million, and Liblib had to partner with Baidu's mobile app.

A Liblib early executive once told us that some investors even told them to "just disband."

Sequoia, Hillhouse, IDG — all looked repeatedly, passed repeatedly. Sequoia had invested in "Toast," founded by Shen Zhenyu, a Liblib competitor; they looked every round, never pulled the trigger, until Lovart emerged; IDG internally pushed Liblib six times, ultimately didn't invest.

Many investors then believed model aggregation platforms were too thin a model, that heavy subsidies and community-building held little value. The favored aesthetic then was big-company senior product managers as founders — people like Leon Ming or Zhang Yueguang — not commercially-oriented Chen Mian.

Chen Mian originally started this venture inspired by CapCut, seeing image generation as commercially large, with user awareness nearly ready, and not requiring head-on confrontation with giants. Simply put: a quiet money-making track.

Until Manus appeared.

Chen Mian has expressed in various settings his shared vision with Manus. Multiple core Liblib members told me that around late 2024, Liblib internally had been conceptualizing a more autonomous form — but hadn't expected it to be an agent.

Many interpret Manus's success as a "timing strategy": at the critical point where underlying model capabilities leap, wrapping them with the fastest engineering into a well-executed concrete product.

Thus, product capability matters less; strategic judgment and execution matter immensely.

On subsequent Lovart, Chen Mian indeed demonstrated opportunistic acumen and Manus-like qualities: also began preparation a year ahead; strong execution — hearing gpt-image-1 would launch in three weeks, immediately pulled the company's best people to Shanghai for closed development; Manus emailed Silicon Valley influencers, Lovart got Elon Musk's like; also cleverly rode the GPT-4o Ghibli trend, even with product imperfections and some hardcoded capabilities, determined to capture "world's first creation agent" mindshare.

In May 2025, Lovart opened beta testing; waitlist exceeded 100,000 in five days — the company's first viral product. Perhaps because it faced overseas markets, Lovart wasn't folded into the current fundraising's Yanyu Group.

But Lovart was indeed the inflection point that turned this company's fortunes.

"After that he understood the rules of this game," said the self-described "lucky" investor who backed Liblib. "Many founders still haven't understood today; some who understood can't execute."

Three months after Lovart's official launch, Sequoia finally had another partner push Liblib through committee, co-leading a $130 million round with CMC and Ant Group. Liblib's valuation leaped from $90 million to $500 million.

An investor at another firm even listed it as his "2024 regret of the year." "If I'd known they were building Lovart, definitely would have invested."

However, some investors still passed even after Lovart. Alibaba, reportedly, didn't bite at this round's $800 million valuation, because Liblib refused to disclose its ESOP.

Mian and Win

Chen Mian's English name is Malvin — reportedly the first half phonetically echoes "Mian," while the second half happens to be "win" in English. Chinese means win (Mian's meaning), English also means win — win upon win.

But in Chen Mian's decade-long career before this, he hadn't experienced true winning.

Born 1992, entered the workforce in 2014, Chen Mian's résumé includes Tencent, 360, Baidu, DiDi, Mobike, Missfresh, ByteDance. He may have worked at the most big tech companies of any AI founder in this wave. But for a long time, he was somewhat ill-fated, joining these famous companies either after core businesses had wound down, or before he was ready to lead.

Until 2020. At ByteDance working on Guaguolong, Chen Mian fought his "most satisfying" battle. Guaguolong's playbook was simple: blanket advertising, lowest prices. But soon, education's "double reduction" policy came. Transferred to CapCut, Chen Mian never entered the core power circle.

Many see Liblib as similar to Genspark: little originality, but with strengths in growth and commercialization. But Chen Mian hates this comparison; the person he loves benchmarking against is — Yiming Zhang.

So much so that there's an internal Liblib saying: one of Chen Mian's drivers is to defeat Dreamina, defeat Kelly Zhang.

You could call this ambition, or partly trauma. His friends, his investors, have received late-night calls — sometimes anxious, sometimes weeping.

During those GPT-4o Ghibli days, one investor received a despondent late-night call from Chen Mian — models could already replicate styles, "Liblib can't survive." Three hours later, Chen Mian called again: no, this makes it even more worth doing!

This chance to win is too precious for Chen Mian.

To this day, most co-founders who followed Chen Mian into entrepreneurship have left him, willingly or not. When they disagreed, Chen Mian would emphasize: "You need to be brain-synced with me." Brain-synced, roughly meaning: think exactly like him.

This March, Roi, who left Liblib to start his own company, announced a funding round calling himself "Liblib co-founder." Chen Mian posted that he had "no true co-founders." This shocked those still at the company. Someone once showed Chen Mian the original investor materials, which clearly listed co-founders' names. Roi was among them.

Chen Mian said (paraphrased): that was then, this is now.

At this company known for high turnover, the shortest tenure I know of is — four days. This former ByteDance 3-2 was recently poached; on day four, finding the work didn't match what was promised, had a major argument with Chen Mian. That afternoon, her Lark account was wiped.

Over three years, Liblib may represent an AI company's growth miracle. But perhaps because it's insufficiently AGI, because of its seeming lack of idealism, it temporarily can't command the most applause.

But the growth Chen Mian believes in may have more silent, devout followers. In an atmosphere of technology-first, innovation-first, Liblib and Chen Mian continue applying a mobile internet success formula; while VCs proclaim they invest in AI-native talent, they direct the most money to Chen Mian.

Multiple founders at smaller-scale companies have told me: AI applications don't really have core technology, models are breathing down your neck, and for Chen Mian to build something this large on such thin ice already commands their respect.

As for after model convergence, perhaps needing better technical teams, perhaps a new game altogether. Who knows? But you have to survive first.

This is also why investors choose them.

At the same time Liblib announced this funding round, I'm told its next round — at $3 billion valuation — is already underway.

Cover image: Paul Delaroche, The Execution of Lady Jane Grey, 1848, Louvre-Lens / Walker Art Gallery