LiblibAI/Evoken: A Vertical AI Image Generation Platform's Hard-Fought Survival in the Cracks of General-Purpose Models

Origin

Chen Mian founded Liblib with the original intention of providing creators with more controllable, more professional AI image generation tools and a community ecosystem when general-purpose model capabilities were still incomplete. The founding team consisted mostly of veterans with hands-on AI product experience; Chen Mian himself had previously fought battles and led teams, and deeply understood the fragility of technology windows. But as Stable Diffusion, Midjourney, and domestic general-purpose models such as Jimeng and Kling rapidly improved, Liblib's core feature differentiation was continuously swallowed directly by model vendors, and the team faced a fundamental existential crisis: 'what we build gets given away for free by the models.' As the feature moat was thinned, cash-flow pressure and an organizational trust crisis overlapped, forcing Chen Mian to rethink whether Liblib was a tool company, a community company, or a middleware company.

Milestones

Turning Points

• General-purpose model capabilities began directly covering Liblib's original core features, forcing the team to abandon the feature-differentiation route and turn to gap scenarios not covered by model vendors
• After the darkest moment of cash-flow crisis and organizational turmoil, the team redefined PMF, cut product lines replaced by general-purpose models, and rebuilt the business model around workflow orchestration and API middleware
• Chen Mian explicitly stated 'not fetishizing niche competition,' shifting from avoiding general-purpose models to becoming a middleware aggregation platform between general-purpose models and users
• Completed $300 million in financing under the overseas brand Evoken at a $2 billion valuation, breaking China's AI application layer financing record, but immediately faced industry questions about whether middleware value is sustainable

Failures & Pitfalls

• In the early days, it bet on feature differentiation, but general-purpose model iteration far exceeded expectations, core features were quickly swallowed, causing both user retention and growth to collapse
• Cash flow once came close to breaking, an internal trust crisis and staff attrition emerged, the founding team went through a 'near-death moment,' and nearly liquidated and exited
• The blind expansion route during the growth phase brought team bloat and feature sprawl, but in reality did not form an effective moat; instead, after the crisis arrived, it was forced to shrink substantially
• During the transformation, there were severe internal disagreements over direction; delayed decisions caused it to miss the best adjustment window, increasing the difficulty of the subsequent turnaround

Core Data

• 估值：Approximately $2 billion (after the 2025 Evoken brand financing) (based on public information, not independently verified)
• 融资额：$300 million (Evoken brand, breaking China's AI application layer financing record) (based on public information, not independently verified)
• 融资历史：Broke China's AI application layer financing record (based on public information, not independently verified)
• 商业模式：Middleware aggregation + API + community + workflow orchestration (based on public information, not independently verified)
• 海外品牌：Evoken (independently operated) (based on public information, not independently verified)
• 团队状态：After the crisis, shrank to the core team, then gradually recovered (based on public information, not independently verified)
• 发展阶段：Post-transformation refinancing stage; cash-flow level has returned to stability, but profit sustainability remains to be verified (based on public information, not independently verified)

Competitors / Peers

Directly benchmarks against Midjourney, Stable Diffusion ecosystem tools, official tool products from domestic model vendors such as Jimeng and Kling, and contemporaneous overseas vertical AI image generation platforms such as SeaArt. It forms a complex 'both dependent and competitive' relationship with general-purpose model vendors—the model resources aggregated by Liblib/Evoken come from these vendors, but their middleware value may be directly swallowed by the vendors' own iterations. After the $2 billion valuation was announced, the industry widely questioned the sustainability of middleware platforms, arguing that the bridge valuation was higher than actual profit support capacity, and that the next round of model iteration could again compress their survival space.