A designer in Hangzhou types a handful of prompts, picks a style model trained by a stranger three provinces away, and ships a product poster in under ten minutes. A year ago, the same job meant a briefing doc, a freelancer, and a three-day wait.
That shift is why a company most Western readers have never heard of just became the clearest signal of where China's generative-media money is flowing.
Key takeaways
- LiblibAI (哩布哩布AI), operated by Beijing Qidian Xingyu Technology, closed a US$130 million Series B in October 2025 — reported as the largest single financing in China's AI application sector that year.
- The round was led by Sequoia China and CMC Capital, with a strategic investor and returning backers (Shunwei, Source Code, Mingshi, Ince) increasing their stakes.
- The platform says it has cultivated more than 20 million AI creators and counts roughly 25 million registered users, according to company-adjacent figures.
- Instead of training its own foundation model, LiblibAI aggregates open and closed models (Qwen Image, FLUX, Seedream, Kling, Hailuo, Vidu, WAN) into one creation workspace — a "tool-plus-community" strategy.
- The bet: in a world where base models converge, the defensible asset is the creator network, the models they publish, and the workflow habits they form.
From a model library to a studio
LiblibAI launched in May 2023, founded by Chen Mian, a former Bytedance product lead who ran commercialization for the Jianying (剪映) and CapCut editing apps. The early product was described as "GitHub for AI painting" — a place where model developers uploaded weights and creators downloaded them.
It has since grown into what the company calls a multimodal creation community, spanning image, video, 3D, and LoRA training. In October 2025 it released LiblibAI 2.0, reframing the product from an aggregator of tools into an "AI professional creation studio" with video generation, multi-model pipelines, and professional effects templates.
The numbers tell the scale story. Beyond the 20 million creators the company cites, it reports over 500,000 user-trained original models and workflows, and more than 500 million images generated on the platform. Those are company-reported figures and should be read as directional, not audited.
Why a community, not a model
The strategic logic is straightforward. As large models (大模型) become commodities controlled by a handful of well-funded players, application-layer companies face a squeeze: the model underneath keeps eating the product on top. LiblibAI's answer is to sit above the models rather than compete with them.
It aggregates both open-source and proprietary image and video generators inside one interface, so a creator can switch from Qwen Image to FLUX to Seedream without leaving the workspace, then move into Kling or Vidu for motion. The platform also runs creator incentives and lightweight model training, building what it describes as a model-scene-creator co-creation loop.
That loop is the moat. A creator who has tuned prompts, trained a custom LoRA, and built an audience on LiblibAI is harder to move than one who simply uses a single model's website.
The compliance floor
LiblibAI passed the Cyberspace Administration of China's deep-synthesis algorithm filing in April 2024 and says it was the first AI community to clear the Generative AI Service Interim Measures备案 in March 2024. It also says it uses blockchain-based provenance and works with the national copyright protection center on watermark standards.
Compliance is not optional in this category. In 2026, state broadcaster CCTV reported that several AI apps, including LiblibAI, had content-safety gaps that let explicit material slip through with oblique prompts — a reminder that community-scale generation carries real moderation risk.
What the money is for
The company says the new capital will go into creator-ecosystem building, R&D, and global expansion. It has already moved outward: an overseas subsidiary launched the design agent Lovart in May 2025, a Chinese-language "Xingliu Agent" followed in July, and the platform announced an international push built on Alibaba Cloud infrastructure in July 2025.
Globalization is the obvious next frontier, but it also means competing with Midjourney, Runway, and a wave of domestic video models for the same paying creators.
Honest limitations
This article relies on company-adjacent and Chinese financial-media reporting; user and creator counts are not independently audited. The strategic investor in the Series B was reported by some outlets as Ant Group and Lenovo Ventures, but the company's own releases name only "a strategic investor," so I have not treated a single name as confirmed. Content-safety scrutiny reported in 2026 is noted but not independently verified here. The generative-media market moves fast, and model-aggregator advantages can erode if base models close their ecosystems.
What readers can do now
- If you create visual content, test an aggregator like LiblibAI before committing to a single model's subscription — the multi-model workspace changes how you compare outputs.
- Creators should export and timestamp their own prompts, parameters, and LoRA configs; provenance features exist, but your local backup is the only enforceable record if a dispute arises.
- Watch the globalization push: a China-built creation studio expanding overseas will shape which models and price points are available to non-Chinese creators over the next year.