HP Teams Up With Peking University-Backed Startup Yuankong Intelligence to Launch an Offline-Operable Model for Scientific Research | Frontline

Text by WANG Xinyi

Recently, Metaspace Intelligence, a team from the School of Science Intelligence of Peking University, released their edge-side AI model Boxer, as well as the AI office platform "Metaspace AI Work" that can be used offline and the scientific research Agent "Metaspace AI for Science". On September 3, the Metaspace Intelligence Edge-side AI Product Launch was held in Shanghai. PANG Dawei, CEO of Metaspace Intelligence, believes that cloud-based large models have some unavoidable practical problems, such as data security, Token cost and stability. To solve these problems, the edge side is a physical solution.

Edge-side intelligence is becoming an unavoidable topic this year. While the intelligence density of small models is increasing rapidly, the capabilities and applications of Agent Harness are increasingly entering daily work, and the computing power of edge-side chips continues to grow.

After accumulating three years of Agent experience, Metaspace Intelligence decided to gradually shift to the edge-side track. Founded in February 2023, Metaspace Intelligence has continuously evolved from initially developing AI Excel office tools and full-link data Agent platforms to launching Claw products, and now its AI office platforms, edge-side intelligence and scientific research Agents.

Boxer is an edge-side model with a size of 35B, which is mainly suitable for office scenarios. In the self-built WorkArena evaluation set of the company, Boxer-35B-A3B-v0-0819 ranked first among models of the same size, showing its advantages in real task capabilities. According to the introduction, this model is post-trained based on the open-source model, and its training data comes from the real task trajectories generated by the Agent Harness accumulated by the team for a long time. Specifically, Boxer can achieve a prefill speed of more than 700 Tokens per second and a decoding speed of more than 30 Tokens per second on a consumer-grade PC with 40TOPS computing power, with running memory occupation of less than 8GB.

At the event site, Dr. NING Kunpeng, Co-founder of Metaspace Intelligence, directly ran three real tasks on a HP computer. Under offline conditions, AI Work calls the local model, which can complete work such as classifying and archiving invoices, checking data anomalies in financial industry transaction flows in just a few minutes, and also performs well in scientific research tasks, which can automatically complete the whole process of complex experiments, drawing charts, writing papers and so on.

At the press conference, Metaspace Intelligence also officially announced its cooperation with HP. For Metaspace, this is to superimpose its own technical capabilities on HP's hardware and commercialization capabilities, pre-install AI Work and edge-side models on HP Z-series models, adopt a standardized delivery, buyout-based business model, and the sales cost on the application side is close to zero. "Every computer in the future may be an AI computer, and the demand for edge-side AI, especially enterprise-level edge-side AI, is about to usher in an explosion," mentioned ZHOU Xinhong, Vice President of China Hewlett-Packard Co., Ltd.

PANG Dawei believes that the current edge-side market is a business that large manufacturers will not engage in for the time being, but he also admits that there is no technology that large manufacturers do not have. The competitive advantage of Metaspace Intelligence lies in the fact that the business model of cloud models makes large manufacturers not fully commit to the edge side for the time being, and the key to making the edge side run - the edge-side environment and data closed loop, is exactly what Metaspace has accumulated for a long time.

The following is part of the exchange record between 36Kr and other media, PANG Dawei (CEO of Metaspace Intelligence), ZHOU Xinhong (Vice President of China Hewlett-Packard Co., Ltd.), YAO Jiayu and Dr. NING Kunpeng (Co-founders of Metaspace Intelligence), slightly edited:

Q: The markets for edge-side AI and cloud-side AI are both very large. How do Metaspace and HP view the edge-side AI market?

PANG Dawei: From the perspective of market size, it is in a period of rapid growth. First of all, the scope that small-size models can adapt to is getting wider and wider, and the size of existing edge-side models is gradually decreasing, from 35B to 27B and even smaller, so more and more local devices can be deployed.

Not only that, from the hardware side, whether it is new devices or old devices, the market for locally deployed AI is also exploding. China ships tens of millions of new PCs every year, and the world stably ships 300 million computers every year, so this market is definite.

ZHOU Xinhong: Every computer in the future will be an AI computer. It can be predicted that the enterprise's expenditure on AI deployment will double in the next two or three years. The explosion of enterprise AI demand is very strong, but this demand has not yet reached the edge side. The topic that AI can assist in office work has only emerged this year. At present, infrastructure, servers and computing power are all in place, and the implementation can be carried out next. Especially for industries with high data compliance requirements such as finance, pharmaceutical companies and manufacturing, data and knowledge must be kept locally, and these tracks have extremely strong explosive power.

Q: At present, the mainstream business model of such office application products is charging by Token. How do you evaluate the buyout and licensing business model of the edge side?

PANG Dawei: We provide standardized delivery, and our business model includes local licensing revenue and cloud token revenue in the edge-cloud hybrid mode. The biggest benefit of cooperating with hardware vendors is that we only need to do a good job in R&D, and we can follow HP's sales system.

In terms of globalization progress, on the one hand, we are developing our own online platform for overseas expansion, on the other hand, we cooperate with HP and follow HP's overseas promotion channels. At present, we are expanding markets in Singapore, Southeast Asia, Japan and South Korea.

Q: Metaspace has experienced many business shifts. After transforming to the edge-side track, what is your positioning?

PANG Dawei: First, we are an edge-side AI model company that develops models, Agents and hardware devices; second, we initially developed ChatExcel, a C-end Chatbot-style office product, which is our core origin. Without cooperation with hardware manufacturers, it is difficult for a small team to realize the self-evolution of models, which is the key for a small team to leverage a large market.

Q: Facewall Intelligence has reached a cooperation with Samsung. Will Metaspace consider cooperating with mobile phone manufacturers?

PANG Dawei: At least at this stage, our choice is definitely to deploy edge-side models on productivity delivery tools, such as PCs and workstations, and we will not enter the mobile phone scenario for the time being. At this stage, mobile phone manufacturers have already pre-installed small models by default. We do not regard mobile phones as target devices, but pay more attention to devices that can truly connect to the network to complete work, form network effects, and scenarios with higher value.

Q: How is your 35B-parameter model trained? Why not directly train vertical industry models, but train general models and mount them on vertical Harness?

NING Kunpeng: This model is post-trained based on the Qwen model using the data generated by our Agent Harness. In terms of effect, the current 35B model has stronger capabilities than GPT-4o more than a year ago, with very high intelligence density. Referring to DeepSeek V4 Flash, which has only 13B activated parameters but already has strong enough capabilities.

YAO Jiayu: As for why we do not directly train industry data into the model, we believe that the real working environment requires the three-way collaboration of models, Agents and devices. Tasks require continuous memory, feedback and operation implementation, which cannot be completed only by a single context or pure chat. Harness is responsible for generating data trajectories, and it must be trained together with the model, the two are inseparable.

Q: Why are the parameters of current models getting larger, but the gains they get are getting smaller?

NING Kunpeng: At present, almost all Internet data has been exhausted, and no more new information can be found. The high-quality data required for the next stage of model training, such as trajectory data, are all in the devices in real use.

Now many large model manufacturers are launching various industry-specific solutions, but enterprises find it difficult to put them into practical use. A phenomenon is that small-parameter models can also achieve good results after proper data proportioning; at the same time, expanding the parameter count by 10 times no longer brings 10 times the gain, and the gap between small models and large models is narrowing.

Q: For users who use edge-side models, how will user data flow back to Metaspace for training?

NING Kunpeng: For users with fully local delivery, the data generated during use will never flow back; for enterprise scenarios, for example, when we provide services for enterprises, enterprises may authorize us to conduct training.

Q: If large manufacturers develop edge-side AI, what are your competitiveness and core barriers?

PANG Dawei: There is no technology that large manufacturers do not have, but each team has its own positioning and business model. The business model of cloud vendors determines that they will not fully commit to the edge side. We assume that all large manufacturers will develop edge-side products, but whether they can do it well and dig deep into the track is another matter. The edge side requires a native edge-side environment and a data closed loop.

Q: HP itself is also developing full-stack AI. Why did you choose to cooperate with Metaspace?

ZHOU Xinhong: AI entrepreneurship includes three parts: data, algorithm and computing power. The data is in the hands of customers, the algorithms and models are in the hands of manufacturers like Metaspace Intelligence, and HP can provide stable computing power and services similar to FDE (Frontline Deployment Engineer).

We have been cooperating with Metaspace for two years, starting from ChatExcel, the first-generation office product of Metaspace AI. Metaspace not only develops the Harness layer, but also has its own trained models, general Agents, and industry-specific Agents corresponding to the finance, pharmaceutical, manufacturing and scientific research industries, which is the differentiated advantage of Metaspace compared with other model manufacturers.

Q: Under what circumstances will the tipping point for the explosion of enterprise-level edge-side AI demand appear?

ZHOU Xinhong: First of all, there is a trend that PC costs will start to rise from 2025, and some manufacturers' server orders have been scheduled to next year, so the price increase will continue. When the price keeps rising, it will inevitably hit a critical point: when manufacturers directly integrate large memory into devices, the cost of a device with local computing power will be almost the same as the money that enterprises spend on cloud services in two or three years, and this critical point will be triggered.

NVIDIA is already doing this, integrating large memory into chips to provide sufficiently strong local computing power. For example, after installing Metaspace's products and models, many tasks can be completed locally, so customers no longer need to continuously pay for cloud APIs. Once such products are launched and users get used to them, the three curves of hardware, software and models will converge, and the triggering of the critical point will be a natural process.

Q: Is the AI wave positive or negative for the PC industry? Lenovo and Dell have taken the lead in forming trends in AI PCs. How will HP make its layout?

ZHOU Xinhong: From the market perspective, the market for PCs priced below $700 is shrinking, but high-unit-price PCs priced above $700 with GPU computing power are seeing double-digit positive growth. Our own sales data shows that the number of units shipped has decreased - because the unit price has increased, the number of units that can be purchased with the same budget is naturally less; but the revenue has achieved double-digit positive growth. The low-price market is no longer our battlefield. The prices of memory, SSD and CPU are all rising, and it is no longer possible to make products at that price point.

Lenovo and Dell are focusing on servers, data centers and cloud-side AI, while HP is purely focused on the edge side, directly entering the edge-side industry, scientific research and enterprise data security markets, and providing enterprises with solutions combined with Metaspace AI to give them more practical and implementable options, which is different from the path of other PC manufacturers.