HyT Capital Portfolio | Mochi Intelligence Completes Over RMB 1 Billion in Angel Series Financing, Valuation Exceeds RMB 7 Billion Within Six Months
This article is sourced from the WeChat public account "暗涌Waves" by 暗涌Waves.
"In the physical world, systems determine life and death."
"暗涌 Waves" has exclusively learned that Mochi Intelligence, an embodied AI company founded only six months ago, has completed over RMB 1 billion in angel series financing within that period. The company is now valued at over RMB 7 billion, making it one of the larger disclosed first-round financings in China's embodied AI sector.
According to "暗涌 Waves," investors in this round include Alibaba, Tencent, BlueRun Ventures, Legend Capital, HyT Capital, Gaorong Ventures, CASSTAR, Source Code Rhythm, Luminous Ventures, Brill Capital, and 58 Industrial Fund.
Mochi Intelligence was co-founded by Huang Qingqiu and Gao Wenli.
CTO Huang Qingqiu, born in 1994, is a quintessential "tech specialist": he earned his bachelor's degree from Tsinghua University's Department of Automation and his Ph.D. from the MMLab at the Chinese University of Hong Kong under the supervision of Lin Dahua. He has long researched robot control, computer vision, and AI algorithms, publishing dozens of top-tier conference papers.
In 2020, he joined Huawei's Intelligent Automotive Solution Business Unit as a "genius youth" and later became head of the autonomous driving AI division, where he led the algorithmic breakthroughs and mass production of Huawei's autonomous driving system from ADS 1.0 to 4.0. He was the first person in the industry to bring a one-stage end-to-end (WEWA architecture) system to million-scale mass production.
CEO Gao Wenli was a co-founder of cross-border logistics company iMile, and was the "key figure" responsible for building out the network across the Middle East and replicating its operating system to global markets. Earlier in his career, he worked at Huawei's carrier business line, leading overseas market expansion for 11 years.
The backgrounds of the two founders determined that Mochi Intelligence would not pursue a single-point demo approach from day one. Instead, it anchors on "hardware-software integrated system engineering capabilities," using commercial service scenarios as its technical training ground, with the ultimate goal of a super terminal for general-purpose home robots.
Embodied intelligence is arguably one of the most complex systems in engineering history. Despite being an industry that also requires tens of billions of dollars in investment, the iteration speed of embodied intelligence lags far behind that of autonomous driving and large language models — because giving a robot a "brain" requires "real operators doing real work in real scenarios" to collect data, and this challenge has not yet been fully solved. But this is precisely the core race in the embodied industry today: a comprehensive competition between general-purpose embodied brains and system engineering capabilities.
Huang Qingqiu believes that without building the brain and without pre-training, this wave of embodied entrepreneurship is meaningless. "With so much capital and talent attracted to this space, if we only focus on the hardware body or single-point tasks, embodied intelligence will struggle to live up to the expectations the industry has today." In his view, the endgame of embodied intelligence is not about whose large model is more "intelligent," but about who can build a mass-producible, reusable, and closed-loop underlying architecture in a chaotic, unstructured physical world.
"暗涌 Waves" met Huang Qingqiu in Shenzhen. We discussed the industry inflection point for embodied intelligence, the endgame judgment for home scenarios, and how autonomous driving experience can empower the embodied sector — which led to this conversation.
The following conversation has been edited by "暗涌 Waves" —
Part 01Building Truly Human-Serving General-Purpose Robots
"Waves": New embodied companies continue to emerge, and some share similar visions to yours. What do you think truly matters?
Huang Qingqiu: If I had to answer in one word, it would be "systems engineering."
In the lab, model architecture sets the ceiling. But in the real physical world at million-unit scale, systems determine life and death.
Today, the industry talks most about VLA vs. world models, end-to-end vs. hierarchical, autoregressive vs. diffusion. But personally, I don't think any of these is the most critical point.
Embodied intelligence, at its core, is about building a hardware-software integrated real-time closed-loop system. What we truly need to do is start from first principles — integrating hardware boundaries, rapid data flow, efficient model iteration, and closed-loop evaluation mechanisms into a single self-evolving organism.
"Waves": What does a real-time closed-loop system mean?
Huang Qingqiu: Real-time and closed-loop are the defining characteristics of Physical AI. Large language models "monologue" in static corpus repositories, while embodied intelligence "fights" in a dynamic physical world.
Take a robot picking up a paper cup filled with water, for example. It has to adjust its force every millisecond based on the cup's weight, the sloshing of the water, and the friction between its fingers and the cup wall — to lift it without crushing it or spilling water. That is "real-time." And every action changes the environment, and the environment change immediately feeds back into the next action — actions and environment are mutually causal. That is "closed-loop."
"Real-time" requires that the operation and motion control models must run on the edge — the brain at least at 10Hz, the cerebellum at least at 100Hz. This frequency threshold pulls models out of the "greenhouse of infinite compute" in the cloud and back into the "cage of power consumption and heat dissipation" on the edge. It constrains compute and caps parameter size — not because we don't want to scale up, but because the laws of physics simply don't allow it. "Closed-loop" means that even a slight loss of sensor precision or a small actuator latency will be amplified by the system, and it also determines what the model learns and how it learns. So at this stage, hardware and software simply cannot be decoupled and iterated independently. The only path to rapid system iteration is hardware-software integration.
"Waves": That is your underlying architectural logic. What about the product side? What is Mochi Intelligence's final product positioning and endgame direction?
Huang Qingqiu: The endgame we see is in the home. Mochi Intelligence aims to build a general-purpose home robot. We will release our first service-oriented robot in July, taking a step forward in that direction.
"Waves": A common question: compared to existing home devices like vacuum cleaners and dishwashers, what is the necessity and core difference of a general-purpose embodied robot entering the home?
Huang Qingqiu: These products are excellent, but they each solve fragmented, point-specific needs. And in solving problems, they often create new "human problems." Users have to rearrange furniture, regularly empty dust bins and change mop pads, load dishes according to the dishwasher's logic, and set up electronic fences for lawnmowers. This is humans serving machines, not machines serving humans.
Robots are the best form to change this status quo. Mochi Intelligence's endgame product will be a home robot with global physical understanding and the ability to handle complex tasks. It will be able to recognize all changes in the home and autonomously perform tasks like cleaning, organizing, and delivering. It won't require you to compromise. It will adapt to your way of life, not the other way around.
Part 02Autonomous Driving Is a Subset of Embodied Intelligence
"Waves": Why are you entering the embodied intelligence space now? Some embodied companies are already actively preparing for IPOs. Don't you feel it's too late?
Huang Qingqiu: Not at all. This is precisely the moment when two technological singularities converge.
The first is the GPT moment. Large language models have excellently solved the problems of high-level semantic understanding and task decomposition. Robots finally have a "cognitive brain." But in my view, a brain alone is not enough. The biggest bottleneck in embodied intelligence today lies in low-dimensional physical execution. The second critical moment is the end-to-end moment in autonomous driving. I personally witnessed this history in the autonomous driving industry — it proved that using neural networks to drive a real-time closed-loop physical entity is feasible.
High-dimensional cognitive intelligence and low-dimensional physical control — these two previously parallel lines are now converging. The point of convergence is embodied intelligence.
As for others preparing for IPOs, that proves that capital is optimistic about this sector in the long term, but it does not mean that the technology or products have already converged. Embodied intelligence is a systems engineering marathon. Getting listed is like getting a supply voucher — the race has only just started.
Besides, we have anchored on the endgame of general-purpose home robots from the very beginning and have taken a hardware-software integrated approach — which is fundamentally different from existing players at the logical level. The home robot market is a trillion-dollar global blue ocean. We have built our company with a global architecture from Day 1. There is no such thing as "entering too late." Perfecting the underlying system and the product is more important than rushing to hit short-term milestones.
"Waves": But many people view autonomous driving and embodied intelligence as two separate tracks.
Coming from the autonomous driving background, how do you see the relationship between the two? And to what extent can the experience from autonomous driving be transferred to embodied intelligence?
Huang Qingqiu: I think autonomous driving is a subset of embodied intelligence. Human operations in the physical world can be summarized as locomotion and manipulation. Autonomous driving deals with two-dimensional planar locomotion problems, while embodied intelligence deals with three-dimensional locomotion and manipulation problems. Moving from 2D to 3D means the solution space explodes — you need to interact with objects and develop a deep understanding of the physical properties of various objects. That is the biggest challenge of embodied intelligence compared to autonomous driving.
As a simplified version of embodied intelligence, autonomous driving offers a wealth of experience to draw upon. The end-to-end technical methodology, the data-loop iteration system, real-time optimization of edge models, engineering solutions for multi-sensor fusion, and even the quality control processes for million-scale mass production — all of these are proven, mature practices that have been validated in the autonomous driving sector. They can all be leveraged in embodied intelligence R&D, significantly shortening the cycle from demo to mass production.
"Waves": What are the commonly underestimated challenges that entrepreneurs in the industry are facing today?
Huang Qingqiu: First, data quality. Future embodied foundation models depend on scalable, high-quality data. How to collect, filter, and process high-quality data is a systemic challenge.
Second, the long-term stability of embodied systems. In unstructured environments, the world is dynamic and full of perturbations. A system that "works" does not mean it can "scale." The engineering gap between shipping 1 unit and 1 million units is immense.
"Waves": Everyone is currently collecting and training data through various approaches. Is this the industry's answer?
Huang Qingqiu: Data is certainly the most important thing, but it depends on what kind of data you are stockpiling. I firmly believe that the endgame form of embodied data must satisfy the "three reals": real operators, in real scenarios, doing real work.
In the endgame, there should be two types of data collection devices. One is purely for capturing video, extremely lightweight — even just a pair of glasses — capturing massive amounts of video for pre-training. The other is an improved version of existing portable devices, finding the sweet spot between wearability and reconstruction precision, capturing relatively smaller quantities of data used for both pre-training and post-training. Both must satisfy the "three reals."
Also, data quality is far more important than quantity. Getting sub-millisecond time synchronization and high-speed motion trajectory accuracy in low-texture environments right — ensuring that every hour of data is high quality — is far more important than rapidly scaling up to tens of millions of hours.
"Waves": What path do you follow on the model side?
Huang Qingqiu: At a simplified level, the model is also a system, essentially doing two things: information compression and modality alignment. The input to an embodied model is high-dimensional tokens with tens of millions of dimensions — images, touch, language, etc. — while the output is just tens of dimensions of action tokens. It is both compression from high dimension to low dimension and alignment across different modalities. With such a low-dimensional output, sparse feedback, and large modality differences, it is very hard to learn.
What engineering needs to do is find ways to provide it with "multimodal dense feedback" from all angles. The various approaches being heatedly debated today are essentially doing this from different dimensions: having the model generate video and predict future frames — that's spatial dense feedback, i.e., world models; having the model express reasoning processes in language — that's logical dense feedback, an important training method for VLA; adding depth maps is geometric dense feedback; adding semantic segmentation is semantic dense feedback.
So world models and VLA are not in conflict at all. It's not an either-or choice. They are "auxiliary lines" laid down for the model from different dimensions. This brings us back to the earlier point: the architecture debate is a false proposition. What Mochi is adopting is precisely a multi-expert architecture that integrates various types of dense feedback, aiming to build a native multimodal model that unifies understanding, generation, and decision-making.
"Waves": Self-developing native multimodal models and pre-training both require sustained heavy investment. How do you evaluate the ROI of this?
Huang Qingqiu: Self-developing a general-purpose embodied brain is indeed a long-term heavy investment, but I don't think this is fundamentally a question about short-term ROI.
If we only stay at the level of hardware bodies or single-point tasks, embodied intelligence will struggle to live up to the expectations that the industry has invested in today. In the long run, the real value will come back to the transferable brain, the data closed loop, and a scalable product system.
Of course, even if we do the math, it still works out. The investment in model training is ultimately a fixed cost. For a high-value product that ships at massive scale, any fixed cost eventually becomes negligible.
Part 03"A Pragmatic Idealist"
"Waves": Was choosing to start a business a difficult decision?
Huang Qingqiu: Starting a great embodied company was a natural choice for me.
When I was at Tsinghua, besides attending classes and sleeping, 80% of my time was spent in Room 508 of the Tsinghua Main Building — Professor Zhao Mingguo's robotics lab. Over those four years, I played with all kinds of robots there, from line-following cars and self-balancing cars, to joining the Vulcan team to build humanoid robots, and finally for my graduation project, I worked with some senior students on a self-driving bicycle that could follow a person. This is where I completed my embodied intelligence enlightenment.
At the same time, I met three senior students who had co-founded SenseTime with Professor Tang Xiaoou and were the first three employees. Through their introduction, I went to SenseTime for an internship. In the early days, SenseTime had just a dozen or so people, all crammed into Wenjin International Apartments in Tsinghua Science Park — the living room was the office, and the bedrooms were the dorms. I still miss that atmosphere of freedom, efficiency, and the pursuit of excellence. That atmosphere infected everyone there. The colleagues who stayed up late coding together back then went on to create many tech star companies like Momenta and MiniMax.
Once you've experienced that kind of fanaticism of building something from 0 to 1, the seed of "building a great tech company" can no longer be contained.
"Waves": During your Ph.D., you worked on "using AI to analyze movies," which sounds completely unrelated to robotics. What impact did that have on you?
Huang Qingqiu: This project gave me early exposure to multimodal unstructured data and trained my ability to build a system from 0 to 1.
At the time, my advisor Lin Dahua told me, "Everyone is researching short videos of a few seconds. Why don't you explore what can be done with movies that are hours long?" So for four years of my Ph.D., I spent most of my time brainstorming various topics on thousands of movies — from actor recognition to multimodal alignment to automated editing.
This was essentially a process of systematically breaking down problems and building small research systems. Building a system is a process of accumulating knowledge over time. There may not be much output before the system takes shape, but once it does, growth becomes explosive.
"Waves": After graduating with your Ph.D., you went to Huawei's Intelligent Automotive Solution BU to work on autonomous driving algorithm R&D. What did your Huawei experience mean to you?
Huang Qingqiu: If my Ph.D. taught me how to "look up at the stars and start new projects," my five years at Huawei taught me how to "keep my feet on the ground and fill the holes" — giving me the ultimate iterative mindset and quality awareness.
Huawei has a quality management system that is nothing short of a "nightmare." Every minor issue is continuously tracked, and before each new version, you face the soul-searching question: "Why hasn't this problem been solved again?" The process was painful, but it was precisely this attitude of not letting any small issue slide and grinding through version after version that produced outstanding products. This accumulation of experience gives me the confidence today to build excellent embodied products and turn ambitious technology into engineering deliverables that never fail in the physical world.
"Waves": If you were to introduce yourself to a stranger without mentioning your resume, how would you do it?
Huang Qingqiu: I'm Huang Qingqiu, a pragmatic idealist.
I firmly believe that within ten years, general-purpose robots will become the "super terminal" for humans in the physical world, entering every household.
But at the same time, I know very well that this day will not arrive suddenly because of a breakthrough paper or a stunning demo. To cross the chasm from demo to product, there is no shortcut — you have to trudge step by step through the mud: grinding through sub-millisecond multi-sensor time synchronization, accounting for every tiny error accumulation in motor actuators, cleaning every trajectory anomaly in the training data, and catching every exception branch caused by loose wiring or packet loss. The road to this super terminal is long and arduous.
Back
Next article
