ZTE and ByteDance’s Nubia M153 Prototype Brings Agentic AI to Life on Smartphones
ZTE and ByteDance just showed off the Nubia M153, a smartphone prototype that runs what they call fully agentic AI right at the operating system level. Unlike typical voice assistants stuck in single apps, this AI sees the phone’s screen, opens apps, taps buttons, fills forms, and handles multi-step tasks on its own—just like a person would. The Indian Express covered a demo thread from ZTE’s CEO that highlights how it books services, makes payments, and even chats with other AI bots.
How the Agentic AI Works
The key is deep integration. The AI, powered by ByteDance’s Doubao model, handles planning and reasoning in the cloud. A separate on-device model called Nebula-GUI from ZTE takes over the screen actions—spotting UI elements, navigating menus, and executing steps fast while keeping sensitive stuff like passwords local, News Mobile reports. This split lets it run smoothly. Doubao supports text and vision, and it’s already got a huge user base in China—175 million people use it, per The Indian Express.
Hardware-wise, it packs Qualcomm’s Snapdragon 8 Elite Gen 5 chipset and 16GB RAM, plus a 6.78-inch screen and 50-megapixel triple cameras, as noted by WinBuzzer.
Real-World Demos That Show Its Power
The demos make the case clear.
- Ask it to find someone to wait in a queue via gig apps—it picks the right service, fills details, and confirms, without the user knowing the app. (Indian Express)
- Snap a photo of a hotel entrance and say you want to book a stay—it IDs the place, checks pet policies on Ctrip, picks dates, and finds the best rate.
- Book a robotaxi, then change the drop-off mid-ride—it reopens the app, updates it, and syncs with the vehicle.
- Order drinks for drone delivery on Meituan—it pays, confirms, and even answers a call from Meituan’s bot to negotiate details autonomously.
- Hunt for the cheapest product online, compare prices across vendors, and order with one voice command—no touches needed except payment confirmation. (NotebookCheck)
- Passive smarts: Snap a pic of something or someone, get context like if a jacket is real police gear or just fashion. Edit photos by voice, like swapping outfits using ByteDance’s image tools.
South China Morning Post points out it books restaurants or edits photos via voice, too.
Why This Matters for Agentic AI
Agentic AI means handing off intent—like “book a trip”—and letting it chain actions across apps without gates or partner limits—a DeepSeek moment 2.0, per Wccftech. Western phones from Apple or Samsung keep AI boxed in approved apps. This prototype changes that by mimicking human use of the GUI. It’s a testbed, with bi-weekly updates planned through March 2026, but already facing pushback: WeChat blocked its login, per NotebookCheck. Still, it shows how OS-level agents could make phones extensions of your brain, especially in China’s app-heavy ecosystem.
Early Demand Signals Big Interest
Priced at 3,499 yuan (about $494), the “engineering prototype” sold out day one online. Buyers now submit requests for production based on demand, and it’s reselling at up to 4,999 yuan on Alibaba’s Xianyu—a 43% markup. (SCMP and NotebookCheck)
China’s edge here comes from blending AI models with everyday hardware fast. No word on global rollout, but this prototype proves agentic AI on phones is real now.
