Li Dahai, Liu Zhiyuan and other co-founders issued an all-staff letter: PB Intelligence will fight three tough battles.
Zhidx August 26 Report: Today, Li Dahai and Liu Zhiyuan, co-founders of Facewall Intelligence, released the all-staff letter "Four Years of Facing the Wall, Breaking Through the General Intelligence Era" on the 4th anniversary of the company's founding, covering key follow-up priorities including edge-side intelligence, industrial scaling, and AI-driven organizational transformation.
"Edge-side intelligence will inevitably become the main battlefield in the general intelligence era." This is a core judgment put forward by Facewall Intelligence in this all-staff letter. Centered on this judgment, the letter reviews the technical routes and product progress of Facewall Intelligence over the past four years, and systematically sorts out three key priorities for the next stage for the first time:
1. Continue to promote edge-side intelligence technology around the "Densing Law"
2. Promote the scaled implementation of existing benchmark cases in vertical industries
3. Use AI to transform internal company collaboration, knowledge management and R&D processes
The letter mentions that Facewall Intelligence is advancing from "validated operation model" to "scale leap". At present, the total global downloads of the open-source MiniCPM series models have exceeded 50 million times, covering text, vision, speech and other modalities; edge-side models have been deployed in automotive smart cockpits, flagship smartphones, as well as scenarios including communities, satellites, legal services and public administration.
In terms of technical routes, Facewall Intelligence judges that intelligent capabilities will continue to evolve along three main lines in the future: perception will move to streaming full-modality, computing will move to higher density, and action will move to autonomous decision-making.
Among them, Facewall Intelligence will continue to take the "Densing Law" as the core technical guideline, reduce the cost of intelligent production and application through architecture innovation, data governance, paradigm evolution and hardware-software coordination, and explore the production mode of "using AI to manufacture AI".
▲ In 2024, Facewall Intelligence and the Tsinghua University team proposed the "Large Model Densing Law", the related paper was published in Nature Machine Intelligence in November 2025, and was selected as the cover article of the current issue.
Meanwhile, Facewall Intelligence mentioned that it will further promote the integration of the model's capabilities of autonomous decision-making, autonomous action and autonomous learning, and build an "edge-side embodied brain" that can independently complete complex tasks.
Industrial implementation is another key priority in the all-staff letter. Facewall Intelligence pointed out that it will cooperate with industry ecological partners in the mode of "edge-side model + agent" to further scale the verified benchmark cases, covering consumer electronics, industrial manufacturing, education, legal services and other fields. The letter summarizes this goal as realizing scalable capabilities that support intelligence deployment for thousands of industries and tens of thousands of terminals.
The third priority points to the company itself. Facewall Intelligence mentioned that it will "polish the company as the most important product", further integrate AI into internal collaboration, knowledge management, R&D processes and decision-making links, and strive to form an edge-side intelligence-driven organizational paradigm.
▲ Facewall Intelligence Beijing Office
The all-staff letter finally reaffirms Facewall Intelligence's judgment on edge-side intelligence. It believes that similar to the division of supercomputing, cloud-side computing and edge-side computing formed in the information age, general intelligence will also form patterns including super intelligence, cloud-side intelligence and edge-side intelligence in the future. This is the core direction emphasized in this letter: to bring general intelligence into more terminals and real scenarios, and finally move towards the "general intelligence era for everyone".
The core viewpoints are summarized as follows:
1. The Densing Law of large models will continue to play a role for a long time in the future. Due to the self-iteration of models, this curve is still accelerating, and its end point is not only "larger models", but also "more efficient models" that reach or even exceed the human intelligence level in the real world with a human-brain-level power consumption of less than 30W.
2. Every physical device may be embedded with a "Jarvis"-style portable agent, which constantly perceives, thinks and makes decisions, while the data never leaves the device.
3. China has the most complete manufacturing system and the largest terminal market in the world, this land is calling for the seeds of intelligence all the time.
4. The intelligence revolution is not a partial capability improvement, but a fundamental leap of intelligence form. We judge that this leap is advancing along three interconnected main lines: perception moves to streaming full-modality, computing moves to higher density, and action moves to autonomous decision-making.
5. Let AI truly integrate into processes and decision-making, so that the organization itself is smarter and more efficient. We will take the lead in verifying the edge-side intelligent organizational paradigm from internal collaboration, knowledge management to R&D processes, and become the organizational paradigm in the artificial intelligence era.
6. What Facewall Intelligence aims to become is not a team that only develops models, nor China's OpenAI or Anthropic; we aim to be the definer and leader of the general intelligence era, and the infrastructure that realizes intelligence for thousands of industries and tens of thousands of terminals.
7. Just as computing power in the information age has been divided into supercomputing, cloud-side computing and edge-side computing, general intelligence will inevitably be divided into super intelligence, cloud-side intelligence and edge-side intelligence.
The full text of the all-staff letter is attached below:
Dear all Facewall members, hello!
It has been exactly four years since we set sail on this journey. First of all, I would like to express my gratitude to every member who has been striving forward with all efforts!
In August four years ago, we started from the Tsinghua University laboratory and founded Facewall Intelligence. At the beginning, like everyone else, we were looking up to the cloud and chasing the scale of parameters; later, we perceived the evolution law of knowledge density of large models earlier than the industry, proposed the Densing Law, and chose a different path: pursuing efficiency, so that intelligence can run into terminals and enter real industries and life scenarios.
Today, four years later, this young scientific research team has grown into one of the leaders in the field of large artificial intelligence models: the total global downloads of the open-source MiniCPM series models have exceeded 50 million times, covering all modalities of text, vision and speech; our edge-side models have been deployed in the smart cockpits of mainstream automotive brands, installed in flagship models of mobile phones around the world, and run on countless terminals and scenarios ranging from communities to satellites, legal services to public administration.
But we want to say that this is only a node in the great journey.
We are advancing from "validated operation model" to "scale leap". This is a critical moment when the technical curve and the industrial curve begin to intersect, and it is also the main battlefield we have been preparing for all the time. The importance of this matter is no less than developing the first operable model in the past, and it is even more exciting. Taking this opportunity, we would like to talk to you about the future we are heading for together.
- What kind of changes we will bring to the world
We believe that the Densing Law of large models will continue to play a role for a long time in the future. Due to the self-iteration of models, this curve is still accelerating, and its end point is not only "larger models", but also "more efficient models" that reach or even exceed the human intelligence level in the real world with a human-brain-level power consumption of less than 30W.
Such a future will profoundly transform the real world we live in:
Every physical device may be embedded with a "Jarvis"-style portable agent, which constantly perceives, thinks and makes decisions, while the data never leaves the device. It is no longer a tool that you "use", but a second brain that grows around you.
The car knows more about your fatigue, emotion and where you want to go next second than you do, even if it loses network signal in the tunnel, it can still make autonomous decisions; on the construction site, tower cranes, drones and ground robots communicate with each other and work in coordination; in the farmland, each crop has its own exclusive agent, which perceives the growth status and pests in real time, and knows when to water, fertilize and harvest; deep in the mine, robots conduct autonomous inspection and independent evaluation, fear no extreme environment, and avoid high-risk operation of humans; the whole factory is alive, capable of self-diagnosis, self-repair and self-evolution...
This is the future that Facewall Intelligence is shaping. It does not float in the cloud, but takes root downwards and grows into the intelligent root system of thousands of industries. This road is not easy: the foundation for a model to be deployed on production lines is the basic model with increasingly higher knowledge density, followed by quantization, power consumption optimization, hardware-software adaptation, and finally numerous "impossible triangles" of good performance, fast speed and low cost are gradually resolved. These tough and tedious tasks are exactly the moat that others are difficult to catch up with, the way we take root downwards, and what we are most proud of. China has the most complete manufacturing system and the largest terminal market in the world, this land is calling for the seeds of intelligence all the time.
This is an opportunity bestowed by the era, and it is also the mission we need to shoulder.
- The three tough battles we need to fight next
First, firmly take the "Densing Law" as the core technical guideline, and expand the technical frontier of human edge-side intelligence.
The intelligence revolution is not a partial capability improvement, but a fundamental leap of intelligence form. We judge that this leap is advancing along three interconnected main lines: perception moves to streaming full-modality, computing moves to higher density, and action moves to autonomous decision-making.
The input of intelligence will inevitably move to streaming full-modality. Intelligence that truly serves people must continuously perceive sound, image, space and environment, understand the physical world people are in and respond in real time. We need to expand the perception boundary of the model, improve the multi-modal understanding and continuous real-time interaction capability, so that intelligence can not only understand human language, but also deeply understand people's situations.
Intelligent computing must be more inclusive and efficient. For intelligence to enter thousands of households and all industries, we need to reduce the cost of continuous use. The "Densing Law" reveals the inevitable trend that the cost of unit intelligence continues to decline. We will firmly grasp the technical methods of architecture innovation, data governance, paradigm evolution and hardware-software coordination, explore the future production mode of "using AI to manufacture AI", so that the cost of production and application of stronger intelligence will continue to decrease.
The output of intelligence will inevitably move to autonomous decision-making, and transform the real world. It is not enough for intelligence to only answer questions and provide suggestions, it also needs to understand goals, take actions, and truly help people solve problems in the real world. We will promote the interconnection of the model's autonomous decision-making, autonomous action and autonomous learning capabilities, build a universal edge-side embodied brain that can independently complete complex tasks, and firmly move towards edge-side intelligent AGI.
The technical frontier does not come by waiting, but by pioneering. We will form our own technical ideas, open up our own technical paths, and lead the industry to realize the final vision of "intelligence covering all things".
Second, thoroughly implement benchmark cases in vertical tracks, and make "intelligence for thousands of industries and tens of thousands of terminals" a scalable capability.
We do not do every single thing in thousands of industries, but choose to be the foundation under these industries. In the mode of "edge-side model + agent", we cooperate with ecological partners who know the industry best, so that intelligence can become a system that can be deployed efficiently, take root in all entities including factories, classrooms, law firms, granaries and aircraft cabins, and integrate into real business flows and data flows. From consumer electronics to industrial manufacturing, from education scenarios to legal ecosystems, we will scale the verified benchmark experience to industrial-level and system-level deployment. This is not just a model, a technology or several companies, but a complete set of new infrastructure.
Third, polish the company as the most important product.
Let AI truly integrate into processes and decision-making, so that the organization itself is smarter and more efficient. We will take the lead in verifying the edge-side intelligent organizational paradigm from internal collaboration, knowledge management to R&D processes, and become the organizational paradigm in the artificial intelligence era.
- A few final words for everyone
What Facewall Intelligence aims to become is not a team that only develops models, nor China's OpenAI or Anthropic; we aim to be the definer and leader of the general intelligence era, and the infrastructure that realizes intelligence for thousands of industries and tens of thousands of terminals. We will use our minds, our hands and our exploration to make the general intelligence era different because of us.
Just as computing power in the information age has been divided into supercomputing, cloud-side computing and edge-side computing, general intelligence will inevitably be divided into super intelligence, cloud-side intelligence and edge-side intelligence; just as the total global edge-side computing power is more than ten times the total cloud-side computing power, edge-side intelligence will inevitably become the main battlefield in the general intelligence era!
In this once-in-a-lifetime golden age, we gather at Facewall Intelligence. The magnificent picture of general intelligence in the next 5-10 years will be written by us together, and this is also our responsibility of the times.
You who are creating the future, do not worry about the long voyage. As long as we forge ahead hand in hand, adhere to the path that is one step ahead of others, fight battles that win more with less, and earnestly do every thing in front of us well, our "persistent efforts will surely steadily achieve our goals". Let's believe, create and achieve together, this will surely become the best memory in our lives.
"Ten years of facing the wall to break through the wall", let's break through the general intelligence era that belongs to everyone together!
Li Dahai, Liu Zhiyuan and all co-founders
August 2026
This article is from WeChat Official Account "Zhidx" (ID: zhidxc om), author: Jiang Yu, authorized for release by 36Kr.