HP teams up with Peking University-affiliated startups to unveil a research-focused model that can run offline | Frontline

Text by Wang Xinyi

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

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

After three years of accumulation of Agent experience, YuanKong Intelligence decided to gradually shift to the edge-side track. Founded in February 2023, YuanKong Intelligence has been evolving from the initial AI Excel office tool and full-link data Agent platform to Claw products, and now to AI office platforms, edge-side intelligence and scientific research Agent.

Boxer is a 35B-parameter edge-side model, which is mainly suitable for office scenarios. In the company's self-built WorkArena evaluation set, 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 pre-filling 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 the running memory occupation less than 8GB.

At the event site, Dr. Ning Kunpeng, co-founder of YuanKong Intelligence, directly ran three real tasks on a HP computer. Under offline conditions, AI Work calls the local model to complete tasks such as classifying and archiving invoices, troubleshooting data anomalies in financial industry transaction flows in 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, YuanKong Intelligence also officially announced its cooperation with HP. For YuanKong, this is to superimpose its 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 and buyout business model, and the sales cost on the application side is close to zero. "In the future, every computer may be an AI computer, and the demand for edge-side AI, especially enterprise-level edge-side AI, is about to explode," said Zhou Xihong, 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 YuanKong Intelligence lies in that the business model of cloud-based models makes large manufacturers not fully invest in 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 YuanKong has accumulated for a long time.

The following is a partial transcript of exchanges between 36Kr and other media outlets with Pang Dawei, CEO of YuanKong Intelligence, Zhou Xihong, Vice President of China Hewlett-Packard Co., Ltd., Yao Jiayu, Co-founder of YuanKong Intelligence, and Dr. Ning Kunpeng, which has been slightly edited:

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

Pang Dawei: In terms of market size, it is in a period of rapid growth. First of all, small-parameter models are compatible with an increasingly wide range of scenarios, 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 support their deployment.

Moreover, from the hardware perspective, the market for local AI deployment on both new and old devices is exploding. Tens of millions of new PCs are shipped in China every year, and 300 million computers are shipped stably worldwide every year, which is a definite huge market.

Zhou Xihong: Every computer in the future will be an AI computer. It is foreseeable that the cost of AI deployment for enterprises will double in the next two to three years. The explosive demand for enterprise AI is very strong, but this demand has not yet been transferred to the edge side. The topic that AI can be applied to office work has only emerged since this year. At present, the infrastructure, servers and computing power are all in place, and the implementation will be realized soon. Especially in industries with high data compliance requirements such as finance, pharmaceutical and manufacturing, data and knowledge must be kept locally, and these tracks have extremely strong explosive potential.

Q: The mainstream business model of such office application products is currently charging by Token. How do you evaluate the buyout and authorization business model of edge-side products?

Pang Dawei: We adopt standardized delivery, and our business model includes local authorization revenue and cloud-side token revenue in the edge-cloud hybrid mode. The biggest benefit of cooperating with hardware manufacturers is that we only need to focus on R&D, and follow HP's existing sales system for product promotion.

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

Q: YuanKong has gone through multiple business transformations. After shifting 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 office product in the form of Chatbot, which is our core advantage. Without cooperation with hardware manufacturers, it is difficult for a small team to realize the self-evolution of models, and this cooperation is the key for a small team to tap into a large market.

Q: FaceUnity has reached a cooperation with Samsung. Will YuanKong consider cooperating with mobile phone manufacturers?

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

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

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

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

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

Ning Kunpeng: At present, almost all public Internet data has been consumed, and no more new valid information can be found. The high-quality data required for the next stage of model training, such as trajectory data, all exist in the devices that are actually in use.

Many large model manufacturers are launching various industry-specific solutions, but it is difficult for enterprises to put them into practical use. A phenomenon is that after the data ratio is properly adjusted, small-parameter models can also achieve good results. At the same time, when the number of parameters expands 10 times, there will no longer be a 10-fold performance gain, and the gap between small models and large models is narrowing.

Q: For users who use edge-side models, how will their data be fed back to YuanKong for training?

Ning Kunpeng: For users with fully local delivery, the data generated during use will never be backflowed. For enterprise scenarios, for example, when we provide services for enterprises, enterprises may authorize us to use their data for training.

Q: If large manufacturers start to 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-side manufacturers determines that they will not fully invest in the edge side. We assume that all large manufacturers will develop edge-side AI, but whether they can do it well and deeply cultivate the field is another matter. The edge side requires native edge-side environment and data closed loop.

Q: HP is also developing full-stack AI on its own. Why do you choose to cooperate with YuanKong?

Zhou Xihong: AI entrepreneurship consists of three parts: data, algorithm and computing power. Data is in the hands of customers, algorithms and models are in the hands of manufacturers like YuanKong Intelligence, while HP can provide stable computing power and services similar to FDE (Frontline Deployment Engineer). We have been cooperating with YuanKong for two years, starting from ChatExcel, the first-generation office product of YuanKong AI. YuanKong not only develops the Harness layer, but also has self-trained models and general Agents, which are divided into different industries, with corresponding Agents for finance, medicine and scientific research. This is the differentiated advantage of YuanKong 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 Xihong: First, let's talk about a trend. The cost of PCs will start to rise from 2025, and manufacturers' server orders have been scheduled for next year, so the price increase will continue. When the price keeps rising, it will definitely reach a critical point: when manufacturers directly embed large memory into devices, the cost of a device with local computing power is almost equal to the money enterprises spend on cloud services in two to three years, and this critical point will be triggered. NVIDIA is already working on this, embedding large memory into chips to provide sufficiently strong local computing power. For example, after installing YuanKong's products and models, many tasks can be completed locally, and 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 critical point will be triggered naturally.

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 Xihong: From the market perspective, the market for PCs priced below $700 is shrinking, while the market for high-unit-price PCs above $700 with GPU computing power is growing at a double-digit positive rate. Our own sales data shows that the number of shipped units is decreasing because the unit price is higher, and the number of units that can be purchased with the same budget is naturally less, but the revenue is growing at a double-digit positive rate. The low-price market is no longer our target market. The prices of memory, SSD and CPU are all rising, and it is impossible to produce products at that price point. Lenovo and Dell are focusing on servers, data centers and cloud-side AI, while HP is fully focused on the edge side, directly entering the edge-side industry, scientific research and enterprise data security markets, and providing enterprises with combined solutions with YuanKong AI to give them more practical and implementable choices, which is different from the paths of other PC manufacturers.