Post-95s Doctor Devotes to World Model, FaceMind Completes Tens of Millions of Yuan Financing

World model company FaceMind recently completed a Pre-A round of financing worth tens of millions of yuan. The investor is Starlink Capital, and the old shareholder 360 made an oversubscribed follow - on investment.

It is reported that FaceMind is advancing its new round of financing. Financial advisors such as Shendu Capital will act as FAs, and some investment institutions have already shown investment intentions.

This is a young AI company. Its helmsman, Lu Hongyuan, a post - 95s generation, founded FaceMind while at school. In the past two years, the company started with end - side full - modality models and gradually shifted to more underlying world models.

As AI enters screens, software, and robots, understanding the world is becoming the next proposition.

A World Model Team Led by a Post - 95s Doctor Emerges

The story of FaceMind begins with Lu Hongyuan.

As a post - 95s founder, Lu Hongyuan studied for his undergraduate and master's degrees at Imperial College London and graduated with a doctorate from the Natural Language Processing Laboratory of The Chinese University of Hong Kong, under the tutelage of Professor Lin Wei. He has long been engaged in research on natural language processing and the underlying mechanisms of large models. During his doctoral studies, he produced 14 first - author/communication papers in top - tier conferences, many of which have become highly cited papers in the field.

In 2023, FaceMind was founded, initially targeting the R & D and application of end - side full - modality models.

What really caught the outside world's attention was the previous discussion about "Ma Jiaqi causing large models to fail". At that time, some large models could accurately state Ma Jiaqi's relevant resume but could not stably output the three characters "Ma Jiaqi". An ordinary person's name unexpectedly exposed the underlying problems of large models in language processing: before text enters the model, it has to be segmented into tokens; when the model encounters low - frequency words, rare person names, or words in small languages, understanding and generation may become unstable.

Lu Hongyuan's team noticed this problem earlier. In 2025, they published a paper related to SLoW, discussing how low - frequency words affect the translation performance of large models. By 2026, their paper result, Adam’s Law, further extended the problem to the sentence level - for the same meaning, the more frequent and common the expression, the easier it is for the model to process and learn.

Even more unexpectedly, the technology related to this paper was adopted by Anthropic, and an Anthropic investor liked and reposted it on X. The judgment of a Chinese post - 95s researcher on the underlying laws of large models was thus seen by more people.

Following this line, FaceMind began to shift its focus to world models.

Simply put, large language models are good at predicting the next piece of text, while world models need to predict what will happen next in an environment. In the context of screens, it means that GUI Agents (Graphical User Interface Agents) understand web pages, documents, buttons, and user intentions; in the field of robots, it means understanding space, actions, and task results.

FaceMind's self - developed world model system is developed around this direction. The company attempts to improve the stability of the model in long - term sequence prediction, screen understanding, and embodied tasks through a cyclically iterative and parameter - efficient model architecture.

Diédie Club is an early verification ground for this set of capabilities. It seems to be an AI danmaku product that can generate interactive danmaku in real - time based on the web pages, documents, videos, or game content that users are browsing. Looking deeper, for a GUI Agent to complete tasks, it must understand the screen, the page structure, judge the button positions, and predict the results after clicking. Each page jump, input feedback, and task completion constitute a type of high - density world model data.

This is also the opportunity that FaceMind wants to seize: world models are becoming the new underlying entry point for AI.

Starlink Capital and 360 Make Moves in the Hottest Embodied Battlefield

The latest financing has come to light.

Recently, FaceMind announced the completion of a Pre - A round of financing worth tens of millions of yuan. This financing not only introduced a new investor, Starlink Capital, but also received an oversubscribed additional investment from the old shareholder 360.

Xiang Qiqi, the pre - investment leader of 360 Group, said, "Dr. Lu is one of the most outstanding young AI researchers I've ever met."

In his view, Lu Hongyuan is not concerned with local optimization but with the underlying principles and architectural innovation of models. When the industry was still discussing the concept of world models, FaceMind had already trained a world model from scratch and achieved industry - leading SOTA - level results in various benchmarking tests. Later, Adam's Law attracted the attention and verification of the overseas leading model manufacturer Anthropic, and the team's newly proposed Loop cyclic architecture further explores the long - term sequence training problem of world models.

"The iteration speed is amazing. Before each communication, I would first look at their latest published papers and technical reports," Xiang Qiqi sighed, truly experiencing what it means to "learn for a lifetime from a single investment."

Li Wenjue, a partner at Starlink Capital, said that the most prominent feature of the FaceMind team is its combination of solid research capabilities and complex engineering implementation capabilities. The core members of the team have long been deeply involved in the underlying technologies of artificial intelligence, able to form independent judgments on cutting - edge directions and quickly verify research results in real - world scenarios.

"We are optimistic about a team with a high density of talents, forward - looking technical judgment, and strong execution ability." In her view, Lu Hongyuan combines the exploration desire of a young researcher and the action ability of an entrepreneur, able to lead the team to continuously tackle difficult problems and translate technical judgments into clear R & D directions. This founder's characteristic and team cohesion are important reasons for Starlink Capital's decision to invest.

In the past year, world models have become a new keyword in the AI industry. Beneath the excitement, differences are also emerging: in the next stage of competition, will it continue to rely on larger data and parameters, or will it improve the model's utilization efficiency of limited data through a new architecture?

FaceMind has chosen the latter.

It is reported that the core features of the company's self - developed model are cyclic iteration and parameter efficiency. Simply put, it attempts to enable the model to have stronger long - term sequence prediction and environmental deduction capabilities with the same parameter scale. The company disclosed that the performance of its 1B - scale model has reached the level of international comparable strong models, and the parameter efficiency has been improved.

Currently, FaceMind has started to verify this set of model capabilities in multiple scenarios. Data shows that its world model capabilities have been verified in simulated embodied environments, GUI Agent environments, and real - machine robotic arm environments. For downstream partners, the company plans to provide a complete set of capabilities, including scenario verification, model training, architecture deployment, inference services, and continuous optimization, for robot body manufacturers, content platforms, chip and cloud manufacturers, etc.

In Lu Hongyuan's view, the opportunities for world models will unfold along with GUI Agents and embodied intelligence. At that time, the competition among models will be about whether they can understand tasks, predict changes, and stably complete actions. After the financing is completed, FaceMind will continue to invest in the R & D of world models and multi - scenario verification.

A young company is squeezing into the card table of next - generation AI infrastructure.