第一财经

WAIC Face-to-Face Subforum: How Qualcomm Builds a Full-Stack Technical Architecture for Agent AI

原文:WAIC面壁分论坛现场:高通如何面向智能体AI打造全栈技术架构

Summary of Key Points

At the 2026 World Artificial Intelligence Conference, Qualcomm emphasized that the era of agent-based AI has arrived, and the prospects for AI applications on edge devices (such as smartphones and headphones) are vast. As agents evolve from single-turn conversations to multi-turn interactions and then to proactive scheduling, the amount of information (tokens) that devices need to process increases by tenfold (from tens of thousands to millions). Therefore, collaboration between the edge and cloud (with the edge handling privacy and real-time tasks, and the cloud handling advanced reasoning) has become crucial. Qualcomm has introduced a new generation of hardware architectures for agents, including Sensor Hub and NPU components, and is working with companies like Mianbi Intelligence to advance edge-side models. In the next 1-2 years, small edge models are expected to achieve the capabilities of large cloud models and will further evolve towards “physical AI,” which can understand the real world.

Detailed Breakdown

1. Edge AI: An Underestimated “Potential Market”

Many people focus on cloud-based AI training (such as large models in data centers), but the edge is actually the “battlefield closest to users.” The global market for terminal devices (smartphones, headphones, cars, etc.) is worth billions, and almost everyone uses them. More importantly, the advent of the agent-based era has sparked a surge in edge-side demand. For example, when using an AI agent on a smartphone to book a flight, it needs to process various pieces of information—such as booking details, checking flight schedules, and selecting seats. The number of tokens required for these tasks has increased dramatically from tens of thousands per single-turn conversation to millions per multi-turn interaction. Relying solely on the cloud for processing would result in high latency and potential privacy breaches. As a result, the edge must take on some of the workload; for instance, sensitive information (like ID numbers) can be processed locally on the phone, while complex tasks like route planning are handled by the cloud. This is what we mean by “edge-cloud collaboration.”

2. Agent-Based Smartphones: From “Obedient Tools” to “Proactive Assistants”

Traditional smartphones simply perform tasks based on user commands (e.g., checking the weather). Agent-based smartphones, on the other hand, can proactively suggest actions based on user preferences. To achieve this, they need three core capabilities:

  • Proactive Environmental Awareness: For example, the phone automatically silences itself when you enter a meeting room or adjusts the seat temperature when you get in the car. This requires a “Sensor Hub” that runs 24/7 with low power consumption, integrating data from cameras, microphones, and positioning systems to provide context information for the AI agent.
  • Powerful AI Computing: The phone’s NPU (AI chip) must be capable of running small models with 2-3 billion parameters (such as miniCPM) to handle generative AI tasks like writing copy or performing translations.
  • Seamless Connection to the Cloud: 5G/6G technology enables seamless switching between edge and cloud computing power. For instance, complex 3D modeling can be handled by the cloud, while real-time voice assistance can be provided locally on the phone.

3. Qualcomm’s “Hardware Solutions” to Power Edge AI

To support agent-based AI, Qualcomm has updated its hardware offerings:

  • Sensor Hub: Collects user behavior data (e.g., frequently visited locations, favorite music) 24/7 to build a “personal knowledge graph” that helps the AI agent better understand the user.
  • NPU Optimization: Uses techniques like model compression and quantization to reduce the size of large models, allowing phones to run models with 20-30 billion parameters smoothly.
  • Newest Chips: The fifth-generation Snapdragon 8 Elite Edition already supports edge AI, and even more powerful platforms will be released this year. Automotive chips will be capable of running large models with tens of billions of parameters, making in-vehicle agents more intelligent.

4. The Future of Edge Models

In the next 1-2 years, small edge models (with tens of billions of parameters) are expected to match the capabilities of large cloud models (with hundreds of billions of parameters). This will be achieved through two main factors: improved hardware performance and algorithmic advancements (e.g., model compression and context-based learning). The long-term goal is to move towards “physical AI,” where devices can not only generate text/images but also understand the real world (e.g., robots recognizing objects or cars planning routes).

5. Collaborative Ecosystems: Working Together to Expand the Market

Qualcomm does not act alone. It collaborates with companies like Mianbi Intelligence to develop edge models (such as miniCPM) and partners to promote cross-device applications (smartphones, cars, robots). For example, in-vehicle agents require a combination of Qualcomm’s chips, Mianbi’s models, and automotive manufacturers’ systems to enable features like automatic seat adjustment. Such ecosystem collaborations can help bring edge AI into everyday life more quickly.

Conclusion

In the era of agent-based AI, the edge is no longer a supporting role; it has become an equal partner to the cloud. Qualcomm’s strategy indicates that future devices (smartphones, cars, etc.) will become increasingly intelligent—capable of not only responding to commands but also proactively solving problems for users. Behind this progress are advancements in hardware, software, algorithms, and ecosystems. For consumers, the most noticeable benefit is that their phones will become more understanding and user-friendly.