Yu Chengdong has started reflecting on the Pangu large model.
```
Author | Huang Yu
Having worked at Huawei for over 30 years, from telecommunications equipment to phones and then to cars, Yu Chengdong is almost always pushed to the forefront whenever Huawei needs a breakthrough.
Now, he has another new task—to bring Huawei's large model back to the center of competition.
On June 12, at the Huawei Developer Conference, Huawei Executive Director, Chairman of the Product Investment Review Committee, and Chairman of Terminal BG Yu Chengdong revealed that on the eve of National Day 2025, the company assigned him to take charge of the Pangu large model.
Regarding the situation of the Pangu large model, Yu Chengdong was straightforward: "When no one in the world knew what a large model was, Huawei released the Pangu large model—it can be said Huawei was an absolute global pioneer in this industry. Later, for various reasons, it wasn't done well, which shouldn't have happened."
As for the future development goals of the Pangu large model, Yu Chengdong once again made bold statements.
He said: "I will lead the team to catch up all the way. In my dictionary, Yu Chengdong, there is no second place, only first place. From being China's number one, to becoming the world's number one in the future."
Looking at the timeline, Huawei does indeed belong to the first batch of domestic enterprises to lay out large models.
Going back to 2021, Yu Chengdong, who was then also CEO of Huawei Cloud, announced the Huawei Cloud Pangu series large models at the Huawei Developer Conference, when almost no one knew about the concept of large models.
However, at that time, Huawei's judgment of AI was more based on the logic of industrial digitalization—the model serves industries first, rather than becoming a consumer-facing product.
This approach was not unreasonable at the time.
But as global competition for large models accelerated since 2023, the industry has seen clear changes. Model capabilities have quickly migrated to terminals, operating systems, and cloud platforms; general capabilities and developer ecosystems are gradually becoming new competitive focuses.
Overseas companies focus their competition on model capabilities, Agents, and system entry points, while domestic internet companies rapidly form rhythms of synergy across models, computing power, and applications.
By comparison, Pangu's presence has always been considerable, but it has not truly established product recognition among the general public. In July 2025, the Pangu large model also fell into a controversy.
Yu Chengdong admitted that becoming number one in the large model field is very challenging. In the AI field, algorithms, computing power, and data are all indispensable, requiring strong engineering capability to support and guarantee.
At this developer conference, Huawei also released the open source Pangu openPangu 2.0.
openPangu 2.0 is equipped with a unified 512K ultra-long context window, and is divided into two versions. 2.0 Pro has a total of 505 billion parameters, with 18 billion active parameters; the lightweight 2.0 Flash has 92 billion parameters, with only 6 billion active, and the sparse ratio significantly reduces the operational load, adapting to deployment needs in various computing power scenarios.
This design philosophy reveals an obvious change—rather than pursuing model size, Huawei has begun emphasizing inference efficiency and engineering capability.
Yu Chengdong also specifically mentioned that some American companies have released models with trillions or tens of trillions of parameters. Why hasn't Huawei done such large-scale models?
First, American companies have more ample computing resources, while Huawei's Ascend computing power mainly supports domestic enterprise needs, leaving very limited cards for itself, and computing power cannot yet meet the requirements for training trillion-parameter large models. Secondly, the computing cost for training large models is extremely high.
At this developer conference, Yu Chengdong also took the opportunity to recruit talent for the Pangu large model, hoping more outstanding talent in the AI field joins the Pangu team.
Doing well with large models is extremely important for both Huawei and Huawei's terminals. For Huawei terminals, one current task is to further enhance the capability of the HarmonyOS. Using AI to restructure the operating system is something HarmonyOS must do well.
In August 2023, Huawei announced that the HarmonyOS 4 system fully connects with the self-developed Pangu large model 3.0, becoming the world's first mobile operating system embedded with large AI model capabilities, empowering the Huawei voice assistant Xiao Yi. In HarmonyOS 5, Huawei further integrated AI capabilities into the system and ecosystem, upgrading Xiao Yi to Xiao Yi Intelligent Agent, capable of executing more complex operations.
2025 is regarded as the founding year of AI Agent, and HarmonyOS 6 also considers Agent as a main focus, aiming to build Harmony intelligence around Agent, and thus has released the new Harmony Intelligent Framework (HMAF), divided into application and agent layers, protocol layer, and platform layer.
Huawei has newly released HarmonyOS 7 this time, announcing that Harmony intelligence will fully evolve toward Agent architecture, targeting Agent's system architecture, Harmony Intelligent Framework 2.0, and the system intelligent agent Xiao Yi as three core upgrades.
Specifically, the new Agent-affinity system architecture has self-evolution capability, integrates terminal-cloud large models, and drives applications toward intelligentization, agentization, and skillification.
The Harmony Intelligent Framework is simultaneously upgraded to 2.0, following the "Intent-as-a-Service" concept. Not only is the access more flexible and development improved, but over 20 system-level AI capabilities are now open, and GUI control capability is opened to developers for the first time.
Meanwhile, Xiao Yi’s ability to execute tasks is further strengthened. Previously, Xiao Yi was more like a voice assistant; now, it begins to take on the role of a system-level intelligent entry point, completing complex tasks through context understanding, cross-app and cross-terminal scheduling, and continuous sensing ability.
The underlying large model capability is crucial for Harmony intelligence. However, in the release of HarmonyOS 7, Yu Chengdong did not disclose which large model was used.
Risk Warning and DisclaimerThe market carries risks, and investment should be cautious. This article does not constitute personal investment advice, nor does it take into account individual users’ special investment objectives, financial conditions, or needs. Users should consider whether any opinions, views, or conclusions in this article fit their particular circumstances. Investing accordingly is at your own risk. ```