Actually, the headline should read: "$5 Billion Valuation, No Revenue, and a 3,800% Download Hype." But that would be too honest for the current bull market narrative.
On July 15, 2025, Mianbi Intelligent officially filed for A-share IPO guidance, joining the queue of Chinese AI startups rushing to capitalize on the Shanghai Stock Exchange's revised fifth set of listing standards. The company, best known for its MiniCPM series of open-source edge AI models, claims over 38 million downloads and integration into mass-produced vehicles from three major automakers: Changan, SAIC, and Geely.
Context: The Edge AI Land Grab
The industry hype cycle has shifted from "bigger is better" to "smaller is smarter." As cloud-based LLMs face latency, cost, and privacy bottlenecks, edge AI—deploying models directly on phones, cars, PCs, and robots—has become the new darling for VCs and regulators alike. Mianbi sits at the center of this pivot, with a narrative that promises "AI on every device, without sending data to the cloud."
But the funding rounds tell a different story. The company raised over 5 billion RMB in the first half of 2025 alone, pushing its valuation past 20 billion RMB (approximately $2.75 billion). Yet the IPO prospectus remains conspicuously absent of revenue, profit, or gross margin figures. The front-runner didn't even bother to fabricate a P/E ratio.
Core: A Systematic Teardown of the Valuation Thesis
Let me be precise: Mianbi's technical achievement is real. The MiniCPM family—ranging from 1.8B to 7B parameters—has been optimized for inference on Qualcomm, MediaTek, and Horizon Robotics chips. The 38 million download count is a legitimate proxy for developer mindshare, not just a vanity metric. However, I've audited enough smart contracts to know that downloads are not revenue, and open-source licenses are not customer contracts.
Here's what the IPO pitch deck won't show you: - The TAM illusion: The edge AI market is projected at $30 billion by 2028, but that includes everything from silicon IP to system integration. The model licensing portion—Mianbi's core business—is a fraction of that, likely under $5 billion. With dozens of competitors (Qwen, Phi-4, Gemma, Llama 3.2), the market is already fragmented. - The revenue invisibility: While the company boasts three automotive OEMs, these are typically single-project contracts for smart cockpit voice assistants. The average contract value is rumored to be under $2 million per model, with no recurring revenue. A bug is just a feature that hasn't been exploited yet? No, a non-recurring revenue model is a feature that hasn't been exposed as a liability. - The regulatory tailwind: The SSE's revised listing standards, announced in June 2025, explicitly allow AI companies with "significant technical capabilities" to list even without profitability. Mianbi's IPO filing came within weeks—a clear sign of policy timing arbitrage, not organic business maturity.
Contrarian: What the Bulls Got Right
To be fair, the bulls have a point on two fronts. First, the open-source community effect is underappreciated: 38 million downloads create a network effect that makes it harder for competitors to displace MiniCPM as the default edge AI framework for Chinese developers. Second, the automotive partnership model—while low-margin—provides a beachhead into the most regulated vertical market. If Mianbi can convert those initial contracts into recurring software subscription services (e.g., over-the-air updates), the revenue model could compound.
But the contrarian angle is also the trap: The company's zero revenue disclosure is not a technical oversight—it's a deliberate signal that the numbers are too ugly to show. In my experience dissecting the 2021 Terra/Luna collapse, when a protocol hides its balance sheet, the fragility is usually exponential.
Takeaway: The Accountability Call
Mianbi Intelligent is a textbook case of regulatory window arbitrage. The company has a real product, a real community, and real traction—but its valuation is built on narrative, not fundamentals. The question is not whether the technology works, but whether the market will sustain 20 billion RMB in valuation before the inevitable revenue reality check. The front-runner didn't disappear; it just moved to the IPO queue.
Data speaks; noise interprets. The 38 million downloads are noise. The three automotive contracts are noise. The only signal that matters is the revenue line—and it's conspicuously absent.
