In the second half of 2026, the competition in the large model track is passing a turning point. As the window for large language models gradually closes, edge computing and AI for Science have become the betting directions for a new batch of entrepreneurs. At this juncture, Pang Dawei, founder and CEO of Yuankong AI, is doing two things simultaneously: embedding models into edge devices for productivity delivery, and using robotic arms and lab benches to drive the dry-wet closed-loop AI for Science.
Pang Dawei's team comes from the School of AI for Science at Peking University. Most of the developers have an interdisciplinary background, having studied chemistry for four years before pursuing a two-year AI program. In March 2023, they developed ChatExcel, one of the earliest AI Excel products in China, transitioning from a cloud application to an edge device over three years. This interview focuses on the future trends of edge AI and the self-evolving closed loop in the field of AI for Science.
Core Perspectives
- The endgame of edge AI is a landscape of vertical players.
The reason opportunities are scattered is that data is dispersed. Capability growth occurs on countless devices, and a single foundation model cannot handle this cycle.
- The edge AI window opens in 2026.
The intelligence density of small models, Harness capabilities, and edge chip computing power—these three curves are maturing simultaneously for the first time.
- Work Is Not Office
Office work is only a small part of Work productivity scenarios. The true value of Work lies in transforming deliverables from documents into business outcomes.
- Scientific research is the scenario with the highest value density at the edge.
When the lab bench becomes the data entry point, competition in scientific research will shift from model capabilities to laboratory assets, and the data generated by the dry-wet closed loop becomes the scarcest barrier.
The following is a transcript of this interview, slightly modified without changing the original meaning.
01. Edge AI Is Naturally a Decentralized Market
Ai Fenxi: Why enter the edge AI market at this specific point in time?
Pang Dawei: We have been accumulating edge technology for several years. We chose to release it this year because three things are maturing simultaneously in edge AI for the first time. First, the intelligence density of models has improved. Today, a 35B parameter model has capabilities close to the full-size models of a year or two ago, and the gains from continuing to stack parameters are significantly diminishing. Second, the computing power of edge chips continues to grow. A consumer-grade computer with 40 TOPS (40 trillion operations per second) can already run models of appropriate parameter scales. Third, the Harness framework for Agents is maturing. Looking at these three curves separately, none are new, but they converge this year, making it possible to run truly productive AI locally.
Ai Fenxi: What is the most direct driving force for doing edge computing? Is it driven by compliance and data sovereignty?
Pang Dawei: Data sovereignty is just one dimension. Uploading sensitive data of enterprises and individuals to the cloud is not a long-term solution. The second dimension is cost. Token costs increase with usage, and tokens are becoming more expensive, which has already entered corporate financial statements. But today, small-parameter models can already do very good things, and costs can be reduced to almost zero. The third dimension is that we believe in a technological trend: edge computing makes intelligence available anytime. Many scenarios cannot all be connected to the cloud. For example, if a long-running task is interrupted halfway, it cannot guarantee continuous and predictable output. This is not a security issue, but a matter of stability and timeliness. These three things determine that localization will become the only way for enterprise-level AI to land.
Ai Fenxi: Besides providing computing power, what other roles does the edge play?
Pang Dawei: Devices are not limited to PCs, nor is it just about placing computing power. In a dark lab, there are robotic arms, petri dishes, and sequencing equipment—dozens of different devices. Driving these devices is a very complex matter, not simply about computing power. For example, putting models and Agents into a robot dog allows video capture to be processed locally. These scenarios that require timeliness all need to be solved on edge devices.
Ai Fenxi: How do you envision the market landscape of edge AI?
Pang Dawei: It will definitely be the coexistence of numerous vertical players, rather than a unified market dominated by a few. The root cause is that edge data is decentralized. What data each device generates and what problems it needs to solve are all different. The types of edge devices are infinite. Besides computers, there are laboratory equipment, LEDs, and even everyday doors and lights. Theoretically, all edge physical devices can be connected to models; it is just a matter of how large a model to adapt and whether it can be scaled. The next step in model capability growth does not rely on how much more static data is mined on the Internet, but on the continuous real interactions happening on devices. Where data is generated, capability grows. Therefore, edge AI is naturally decentralized, and a unified platform cannot handle it.
Ai Fenxi: With every device being different and every scenario being different, won't it become too fragmented?
Pang Dawei: Each team has limited resources and endowments, so we only focus on two scenarios: Work and scientific research. There are many opportunities in industry, but we will not focus on them for the time being.
Ai Fenxi: What are the opportunities for large model companies at the edge?
Pang Dawei: The window period for large companies to achieve unification has ended. What we do at the edge is to get into the devices. This must be a new idea, and what we do in scientific research is also a new idea. It is not that new means right, but we need to be the pearl of the next stage. Otherwise, everyone is doing something repetitive that will inevitably be done by large companies, which is meaningless.
02. AI Work Is Not Office; Deliverables Must Break Out of the Three-Piece Suite
Ai Fenxi: The office market is not very large. Why are large companies investing all their efforts? How to better understand AI Work?
Pang Dawei: As long as it is work, it belongs to Work. Work is not just about doing office tasks or using Office software; collecting and analyzing product information on shelves is also Work. Office software has been around for so many years, and the market size is just so big. Why is everyone still doing it? Actually, everyone will gradually realize that Work is actually a productivity scenario, not just Office.
Ai Fenxi: For example, Yuankong AI, what can be delivered in edge AI Work?
Pang Dawei: Our product is positioned as an omni-scenario intelligent productivity platform, covering Excel, Word, PPT, databases, as well as code and files, handling desktop work scenarios. What we deliver are results, not documents. We demonstrated three things at our launch event: organizing invoices, checking abnormal transactions, and running the entire scientific research process. The output of none of these is a document, but a business result. Users only need to clarify their goals in natural language, and the rest is executed by the system itself. It supports 24-hour continuous task execution and can resume from checkpoints after interruptions.
Ai Fenxi: Yuankong AI has post-trained two models, 35B and 27B. What is the starting point for choosing these two parameter points?
Pang Dawei: One is a 35B MoE model, and the other is a 27B dense model. The Work scenarios applicable to the two are still different. If you want to get things done, you must be equipped with a very good Agent framework. The effect of an Agent framework driven by models below 20B is discounted.
Ai Fenxi: What is the difference between the business model of edge AI Work and the cloud?
Pang Dawei: The biggest problem with the cloud is that the more users you have, the more you lose. We are exactly the opposite; the more users we have, the more profit we make. We emphasize standardized delivery. For example, entering HP's devices, we make money for every unit shipped. We almost only need to bear R&D costs, and sales costs are almost zero.
Ai Fenxi: How do you view large companies currently pushing AI Work with full force?
Pang Dawei: Large companies always want to pursue universality, but when entering the enterprise side, the problems encountered are still the old ones: how much can you deliver, can you make money, and is there profit.
03. Scientific Research Is the Highest-Value Validation Field for Edge Computing
Ai Fenxi: Why is scientific research the preferred scenario for the edge?
Pang Dawei: Scientific research has a high value density and strong requirements for localization. Scientific research itself is confidential. Models need to be stored in separate devices, and the process must be traceable, reproducible, and verifiable, so edge computing is inevitable. But edge computing is just a technical implementation; it does not determine whether scientific research can succeed. The real threshold lies in data and closed loops.
Ai Fenxi: How to understand this?
Pang Dawei: In the pre-training stage, static data on the Internet has been mostly consumed, and the benefits of continuing to stack data are constantly declining. What is truly scarce is the trajectory data left by models working in real environments: why it was done, how it was done, and whether it was finally accomplished. These data will not appear on the Internet; they only appear on devices. Therefore, we emphasize that the RSI (self-evolution) in scientific research must be a dry-wet closed loop. Without a dry-wet closed loop, it cannot be called RSI. In this closed loop, the dry part involves running algorithms, making hypotheses, and doing simulations; the wet part involves using Agents to drive all devices to execute the entire process. Only through the operation of the lab bench can you obtain all the data and feed it back for the self-training of the model.
Ai Fenxi: So models, Agents, and devices must be bound together for the edge?
Pang Dawei: Yes. Relying solely on models is not enough, and relying solely on Agents is not enough either. They must be combined; models and Agents are like the left and right legs. Devices are not just carriers of AI capabilities, but also the starting point for the next round of capability growth. Give an Agent a task, it goes to work, and after finishing the work, it feeds the data back so the model can continue to evolve.
Ai Fenxi: In the field of scientific research, when the lab bench becomes the data entry point, will the competition shift from the triple competition of models, Agents, and devices to the competition for occupying laboratory assets?
Pang Dawei: To be precise, the real data in laboratories is becoming increasingly valuable. For example, in the research of a certain small molecule, there may be only five laboratories in the world doing it, and once it is done, it is done. Other companies cannot do it even if they want to, because you don't even have the lab bench data.
Ai Fenxi: In the field of scientific research, do completely different models need to be trained?
Pang Dawei: Yes, there will be different models for scientific research. For example, those working on proteins, small molecules, materials, and atmospheric physics are completely different models. Therefore, scientific research involves models of completely different scales and modalities. Not only the models, but the equipment is also different. Atmospheric physics requires sensor acquisition equipment, while protein research requires sequencing equipment.
Ai Fenxi: How is the progress in the field of scientific research?
Pang Dawei: In 2027, we will strive to make real scientific research outputs in several of our pipelines. We have already cooperated with several top labs and universities and are currently working on it. The result we want is that the wet lab benches in several cooperating pipelines can truly produce scientific research results. It is not about publishing papers; publishing papers is easy. It is about truly producing results. Achieving self-evolution of scientific research models is the goal we pursue.