Woofun AI reports that X Corp's FaceMind has finalized a Pre-A financing round valued at tens of millions of yuan, with Starlink Capital leading the investment while existing shareholder 360 participated in the additional funding. The company is currently advancing toward a subsequent financing round, facilitated by financial advisors including Shendu Capital, as multiple investment institutions have already signaled strong interest in participating. This capital injection underscores the market's growing conviction in the transition from standard large language models to more complex world models capable of environmental prediction.

FaceMind operates as a young artificial intelligence enterprise founded by Lu Hongyuan, a post-95s entrepreneur who launched the venture while still pursuing his academic studies. Over the preceding two years, the organization concentrated its resources on developing end-side multimodal models before strategically pivoting toward the more foundational challenge of world models. As artificial intelligence systems increasingly penetrate screens, software interfaces, and robotic hardware, the capacity to fundamentally understand the physical and digital world has emerged as the critical next hurdle for the industry. Lu Hongyuan's academic background includes bachelor's and master's degrees from Imperial College London, followed by a doctorate from the Natural Language Processing Laboratory at the Chinese University of Hong Kong under the supervision of Professor Lin Wei. During his doctoral tenure, he published 14 research papers, several of which achieved high citation rates within the field of natural language processing and large model mechanisms.

The company was formally established in 2023 with an initial mandate to develop and apply end-side multimodal models, but public attention was sharply drawn to a specific incident involving the inability of a large model to accurately generate the name "Ma Jiaqi." Although the model could retrieve relevant information regarding the individual, it failed to correctly spell the name, exposing a structural flaw in how large models process language through tokenization. When text is fed into a model, it must first be fragmented into tokens, a process that causes instability when the model encounters low-frequency words, rare names, or terms from lesser-known languages. Lu Hongyuan's team identified this vulnerability early, publishing a paper in 2025 on SLoW that analyzed the impact of low-frequency words on the translation performance of large models. By 2026, their research evolved into the concept of Adam's Law, which explored these issues at the sentence level, positing that frequent and common expressions are inherently easier for models to process and learn.

Notably, the technology detailed in this paper was subsequently adopted by Anthropic, with one of Anthropic's investors sharing the findings on X Corp, thereby bringing the insights of this young Chinese researcher to a global audience.

Building upon this theoretical foundation, FaceMind initiated a strategic shift toward world models, which differ fundamentally from large language models that excel at predicting the next segment of text. World models aim to predict future events within a given environment, a capability essential for GUI Agents to understand web pages, documents, buttons, and user intentions, as well as for robotics to comprehend space, actions, and task outcomes. FaceMind's self-developed world model system is engineered around this objective, striving to enhance model stability in long-term prediction, screen understanding, and embodied tasks through iterative approaches and efficient parameter architectures. DianDianShe serves as an early testing ground for these capabilities, appearing superficially as an AI-based bullet-screen product that generates interactive comments in real-time based on user viewing content.

However, at a deeper operational level, the system requires a GUI Agent to understand screen layouts, determine button locations, and predict interaction results, generating high-density data for the world model with every page transition and task completion.

Woofun AI data shows that the latest funding round not only introduced Starlink Capital but also secured additional capital from 360, validating the team's approach to core model innovation. Qi Xiangqi, who oversees investment efforts at 360, stated that "Dr. Lu is one of the top young AI researchers I have ever met," emphasizing that Lu Hongyuan focuses on innovative principles and architectures rather than partial optimizations. While the broader industry was still debating the theoretical viability of world models, FaceMind had already commenced training world models from scratch, achieving industry-leading results across various benchmark tests. The team's Loop architecture further explored challenges related to the long-term training of world models, reinforcing the value of their research trajectory. Li Wenjue, a partner at Starlink Capital, highlighted that the team's most outstanding feature is the combination of solid research capabilities with the ability to implement complex technological solutions in practical applications. She noted that core members have long dedicated themselves to fundamental AI technologies, forming independent judgments in cutting-edge areas while rapidly verifying research findings in real-world scenarios.

The investment community values a team possessing high talent density, forward-looking technical judgment, and strong execution capabilities, qualities that Li Wenjue identified as key drivers for Starlink Capital's decision to invest. In her assessment, Lu Hongyuan combines the exploratory spirit of a young researcher with the action-oriented mindset of an entrepreneur, enabling him to lead the team in tackling challenging problems and translating technical insights into clear research directions. Over the past year, the concept of world models has become a dominant buzzword in the AI sector, yet significant disagreements persist regarding the direction of future competition. The central debate focuses on whether the industry should continue to rely on larger datasets and increased parameters or shift toward improving model efficiency in handling limited data through new architectures. FaceMind has explicitly chosen the latter path, asserting that the core features of its self-developed models are iterative design and efficient parameter utilization.

The company aims to enable models to achieve superior long-term prediction and environmental inference capabilities using the same amount of parameters as competitors. According to FaceMind, its 1B-scale models have reached performance levels comparable to internationally renowned models while demonstrating significantly improved parameter efficiency. Currently, the company is testing these model capabilities across diverse scenarios, including simulated embodied environments, GUI Agent environments, and real robotic arm systems. Moving forward, FaceMind plans to provide partners in the robotics industry, content platforms, and chip and cloud technology companies with a comprehensive service suite ranging from scenario validation and model training to architecture deployment, inference services, and ongoing optimization. Lu Hongyuan believes that opportunities associated with world models will expand in tandem with the development of GUI Agents and embodied intelligence, where the true test of these models will be their ability to understand tasks, predict changes, and perform actions stably. This marks a definitive shift as the company moves from theoretical research to becoming a foundational infrastructure provider for the next generation of artificial intelligence.