With 20 billion yuan quietly stockpiled, how thick is the competitive moat?

On July 15, according to media reports, Facewall Intelligence completed a new round of financing. In the first half of 2026, its total cumulative financing exceeded 50 billion yuan, with a valuation breaking through 200 billion yuan, making it the unicorn with the highest valuation in the end-side intelligent track. Four days later, at the WAIC 2026 venue, Facewall Intelligence announced another set of figures: the total global downloads of the MiniCPM series open-source models have exceeded 38 million times.

A company that grew out of a Tsinghua University lab has turned a "non-consensus" business into a "consensus" that the capital market is willing to pay 200 billion yuan for in just four years.

However, when the capital consensus is formed, it is usually the beginning of the dilution of the first-mover advantage. Over the past two years, giants such as Alibaba, ByteDance, Apple, and Qualcomm have intensively entered the end-side market. What Facewall Intelligence needs to prove next is no longer just whether the models can run on end devices, but whether these technologies can generate sustainable revenue and solidify into real commercial barriers.

01.

50 Billion in Half a Year: Who Is Placing Their Bets

The financing history of Facewall Intelligence follows a clear main line.

According to reports, the angel round of Facewall Intelligence was led by Zhihu. In June 2023, Li Dahai, then CTO of Zhihu, took up the position of director and CEO of Facewall Intelligence, responsible for the company's strategy and daily operations. In 2024, a new round of financing was led by Huawei Hubble and Chunhua Venture Capital, with follow-on investments from institutions such as the Beijing Artificial Intelligence Industry Investment Fund.

In May 2025, Hongtai Fund, Guozhong Capital, Tsinghua Control Jinxin, and Moutai Fund made a joint investment. By December of that year, Jing Guorui, Cas Investment, CICC Porsche Fund, Miju Capital, and Heji Investment jointly injected several hundred million yuan.

The real acceleration happened in 2026.

At the end of February, Facewall Intelligence completed the first round of financing after the Spring Festival, led by China Telecom, with follow-on investments from CITIC Goldstone and CITIC Private Equity. In April, a new round of financing was jointly led by Shenzhen Venture Capital and Inovance Industrial Investment, with follow-on investments from institutions such as Daohe Long-term Investment, Guotai Junan Innovation Investment, and Wuyuefeng Science & Innovation. By then, state-owned capital from Beijing and Shenzhen, as well as multiple industrial capitals, had all joined the shareholder list.

The new round of financing disclosed in July has also introduced investors such as national-level funds, central state-owned enterprises, and automobile manufacturers. According to Facewall Intelligence, the company's total cumulative financing in the first half of the year has exceeded 50 billion yuan.

In this shareholder list, there are almost no internet giants like Alibaba, Tencent, and ByteDance that have laid out cloud-side models, end-side hardware, and user entrances at the same time. In contrast, large model companies such as Zhipu AI and MiniMax have received capital support from major manufacturers to varying degrees. The capital structure of Facewall Intelligence is more inclined to "state-owned capital + industrial capital": it includes not only telecom operators and automobile manufacturers, but also local state-owned funds and industrial automation enterprises.

The underlying business logic is not complicated. Facewall Intelligence hopes to become a model supplier for car companies, mobile phone manufacturers, and robotics companies, and maintaining relative neutrality is conducive to expanding its customer base. If it is deeply tied to a single internet giant, it may instead affect its cooperation with other ecosystems.

Capital market policies have also changed investors' exit expectations. In June 2026, Wu Qing, Chairman of the China Securities Regulatory Commission, announced at the Lujiazui Forum that the scope of application of the fifth set of STAR Market listing standards will be expanded to the artificial intelligence sector, to support high-quality large AI model enterprises to go public.

This does not mean that Facewall Intelligence has already obtained the admission ticket for listing, but for large model companies that have not yet generated stable profits, the domestic listing channel is becoming clearer. For the state-owned capital and industrial capital that entered in this round, Facewall Intelligence is no longer just a technology project that requires long-term waiting, but may also become an investment with a clear exit direction.

02.

"Densing Law": Academic Achievement, Also a Financing Narrative

The core pillar that allows Facewall Intelligence to tell the story that "we understood the end-side earlier" is the "Densing Law" proposed by Liu Zhiyuan's team.

The "capability density" referred to in this law is the ratio of the effective parameter scale of the model to its actual parameter scale. To put it simply, it means using fewer parameters and lower inference costs to achieve tasks that previously could only be completed by larger models.

The related paper was published as a cover paper in *Nature Machine Intelligence* in November 2025. The paper analyzed 51 open-source foundational models and fitted an empirical curve across five commonly used benchmarks: the maximum capability density of open-source models roughly doubles every 3.5 months.

The rise in Facewall Intelligence's international popularity is also related to a plagiarism controversy.

In May 2024, a team composed of two Stanford undergraduates and another researcher released Llama3-V, claiming to have trained a high-performance multimodal model for only about 500 US dollars. The project was soon questioned for extensively reusing the MiniCPM-Llama3-V 2.5 jointly released by Tsinghua University and Facewall Intelligence. In addition to the highly similar model structure and configuration, the two models even made identical errors when identifying undisclosed Tsinghua bamboo slip data.

The two team members subsequently issued a public apology and took down the project. Christopher Manning, director of the Stanford Artificial Intelligence Laboratory, criticized them for not facing up to their mistakes; while Lucas Beyer, then a researcher at Google DeepMind, lamented that the MiniCPM with comparable performance had received far too little attention previously.

The first time Facewall Intelligence attracted widespread attention in the international open-source community was not through a press conference, but through the misstep of its competitor.

However, the cycle of "density doubling" has shown obvious changes at different stages.

In multiple speeches and interviews in mid-2024, Liu Zhiyuan and Li Dahai used the statement that "knowledge density doubles on average every 8 months". By the end of 2024, when the "Densing Law" was formally proposed, the statement had changed to roughly every 100 days, that is, doubling every 3.3 months. The final results presented in the paper published in *Nature Machine Intelligence* are approximately 3.5 months.

Li Dahai, CEO of Facewall Intelligence

These sets of figures are not entirely comparable. The early statements are closer to conceptual observations, while the formal paper defines the model samples, time range, and evaluation methods. But this also means that the "Densing Law" is more appropriately understood as an empirical curve that needs to be continuously updated, rather than a natural law that can stably predict the situation in the coming years.

The paper itself also acknowledges that current measurements of model capability still rely on limited benchmark tests, which may be affected by factors such as data contamination and evaluation set saturation, and capability density cannot grow indefinitely.

Facewall Intelligence's claim that end-side models have "caught up with the GPT-4 era" also needs to be viewed with caution. During the 2026 Zhiyuan Conference, Li Dahai cited the AA-Index ranking, stating that the score of MiniCPM5-1B is already close to that of GPT-4o in 2024.

This comparison can illustrate that the efficiency of small models is improving rapidly, but a single ranking cannot cover the comprehensive performance of different models in reasoning, coding, multimodal processing, and real-world tasks. Summarizing it as "the intelligence level has caught up with GPT-4" has excellent communication effects, but it is not yet a technical conclusion that has undergone unified benchmarking and third-party cross-validation.

03.

Three Commercial Pipelines Have Not Disclosed Revenue Figures

Facewall Intelligence's current commercialization mainly relies on the automotive, mobile phone, and enterprise-level markets, while embodied intelligence is a newly added business line.

Automotive is a business with relatively clear implementation paths. In 2025, Facewall Intelligence's end-side models were already deployed in mass-produced models such as the Changan Mazda EZ-60 and Geely Galaxy M9. The company also signed a strategic cooperation framework agreement with Aptiv, hoping to enter more markets through the latter's automotive supply chain.

However, the statement in some previous reports that "more than 300,000 mass-produced cars have been delivered" is inaccurate. The official statement given by Facewall Intelligence is that it expects 300,000 cars to be equipped with its end-side models by the end of 2026. This is an annual target, not the number of deliveries that have already been completed.

In the enterprise-level market, Facewall Intelligence claims that its agent platform PilotDeck has achieved large-scale commercial use, and during WAIC, it launched CPM for Legal, which is tailored for professional legal services. However, existing public materials still do not disclose specific customer quantities, contract amounts, and revenue contributions.

On the mobile phone side, Facewall Intelligence officially announced that the MiniCPM series will be deployed on several flagship Samsung models. The cooperation between the two parties has entered the product development and filing stage, but the specific models, shipment schedules, and how much revenue Facewall Intelligence can obtain from them are still unclear.

Embodied intelligence is the latest business line. On July 19, during WAIC, Facewall Intelligence released the MiniCPM-Robot series, including the general vision-language-action model MiniCPM-RobotManip and the mobile tracking model MiniCPM-RobotTrack. Among them, MiniCPM-RobotManip has 1.5 billion parameters and can complete long-horizon operation tasks such as making sandwiches.

In the RMBench evaluation results disclosed by Facewall Intelligence, MiniCPM-RobotManip scored about 53 points, while π0.5 scored about 10 points. However, RMBench mainly examines the performance of robots in operation tasks that depend on contextual memory, and cannot represent the comprehensive capabilities of the model in all robotic scenarios.

These figures seem considerable when viewed in isolation, but most of them are download volumes, planned deployment quantities, cooperation counts, and evaluation scores. Facewall Intelligence has never publicly disclosed its actual revenue, gross profit margin, and customer renewal status.

This is not uncommon among unlisted AI companies. But for a company that has raised more than 50 billion yuan in half a year and has a valuation exceeding 200 billion yuan, the absence of commercial data still deserves attention. Especially after the listing channel has become gradually clear, what the public market ultimately cares about is not the number of model downloads, but the quality of revenue and cash flow.

Facewall Intelligence's choices in the C-end market are also notable. The conversational assistant "Facewall Luca" was launched in May 2023 and opened to the public in November of the same year, after which continuously updated user data was rarely seen in public channels. Currently, Doubao has reached 382 million monthly active users, and Qwen has also reached 167 million monthly active users.

Facewall Intelligence did not continue to invest heavily in the C-end, mainly due to resource trade-offs. It has placed its main bets on supplying foundational models to car companies, mobile phone manufacturers, and robotics companies. This path avoids the expensive user acquisition and operational costs, but it also deepens the company's dependence on a small number of large customers, hardware cycles, and project deliveries.

At least based on public information, Facewall Intelligence has not yet formed an independent revenue source that can directly reach consumers and hedge against its To B business.

04.

The End-Side Enters the Red Sea, and Moats Shift to Chips and Systems

Facewall Intelligence is indeed one of the earlier large model startups in China that took end-side models as its main line. The MiniCPM released in February 2024 also helped it establish strong recognition in lightweight models and end-side deployment.

But on a global scale, the situation is different. Google released Gemini Nano, which runs on mobile chips, as early as December 2023. Apple then announced its end-side foundational model with about 3 billion parameters in June 2024.

At WAIC 2026, the Beijing Academy of Artificial Intelligence, in conjunction with the Beijing Key Laboratory of End-Side Intelligence, released a report defining 2026 as the "first year of large-scale implementation" for end-side intelligence.

According to the forecast of Frost & Sullivan, the global end-side AI market size will increase from 3.219 trillion yuan in 2025 to 12.2 trillion yuan in 2029, with a compound annual growth rate of approximately 39.6%. Policy signals have also become clearer: the 2026 government work report first proposed to "foster new forms of the intelligent economy". The previously issued "Opinions on Deeply Implementing the 'AI+' Action" by the State Council proposed that by 2027, the penetration rate of applications such as new-generation intelligent end devices and agents should exceed 70%.

The problem lies precisely here. After end-side AI has evolved from a relatively niche path into a widespread consensus, the players at the table have become completely different.

Alibaba continues to open-source Qwen models of various sizes and integrate them into hardware such as the Quark AI glasses; ByteDance is still adjusting the product plans for its Doubao AI glasses; Apple launched the Core AI framework for Apple chips in 2026; Qualcomm, MediaTek, and Intel are also building their respective end-side AI platforms. Mobile phone manufacturers are constantly embedding large model and agent capabilities into their operating systems.

These companies respectively control cloud-side models, chips