Podcast Atlas
concept Updated 2026-07-09 Tags: Ai, Edge-Ai, Hardware, Model-Architecture

On-Device Model Hierarchy

On-Device Model Hierarchy is Yang Meng / 杨萌’s distributed-intelligence frame in 144. 对杨萌的4小时访谈：消费电子死与生、第三类公司、端侧模型、产品方法、游戏模式. He expects future AI systems to combine trillion-scale cloud “brains,” billion- to ten-billion-scale endpoint models, and million- to ten-million-scale perception/control models inside sensors, actuators, and small devices.

The analogy is biological: a human body does not send every signal to one central super-brain. Eyes, ears, muscles, and other organs perform local processing while higher-level planning happens elsewhere. In product terms, the model size should match the problem’s complexity, latency, privacy, and power constraints.

AI 时代的超级入口还是手机吗？｜ S10E17 adds the smartphone implementation version. Han Boxiao says vivo adapts smaller models to phone platforms and special foldable system versions, while Chen Yiqiang explains why MediaTek needs NPU architecture, middleware, and scheduling so endpoint models can run efficiently inside everyday phone workloads.

Key Claims

  • Hardware AI should not be reduced to one giant model controlling every device from the cloud.
  • Different tasks need different model scales: speech separation, local video search, robot control, and general planning do not have the same compute requirements.
  • Endpoint and local models can improve privacy, latency, and reliability when continuous cloud access is undesirable.
  • The hierarchy makes hardware more than a model-company accessory; devices become distributed organs of perception and control.
  • It also changes product design: the user cares about smoother experience, not whether intelligence came from a cloud model, base station, chip, or tiny local model.
  • The smartphone version adds a system constraint: local models must share battery, heat, memory, display, CPU, GPU, and NPU budgets with the rest of the phone experience.

Connections