M50 Inside: Engineering 100B-Parameter Edge Inference or Just a New Power Wall Game?

At the Zhangjiang exhibition area of WAIC 2026, I saw Houmo Intelligence's demo with my own eyes: a palm-sized device driving a hundred-billion-parameter large model, generating PPTs in real-time, and powering a holographic digital human for continuous conversation. While people at the neighboring booth were still discussing the API call costs of cloud-based large models, this team had already squeezed the model into a local terminal. As an engineer who worked on 5G baseband power optimization at Unisoc, my first reaction wasn't excitement, but skepticism—has the power consumption of this "palm-sized" box been tested? What's the thermal dissipation solution? How is the mass production yield?