STEPX Neo at WAIC 2026: When a Phone Stops Being Just a Phone

At WAIC 2026 in Shanghai, one of the most crowded booths belonged to a phone. Not a foldable, not a five-camera monster, not a benchmark champion. It was STEPX Neo, built by StepFun—and it walked away with the conference's "Treasure of the Exhibition" award.

Three identical STEPX Neo units sat on the booth. Each behaved differently: one was reserved, one chatty, one intimately familiar with its user's social life. The explanation from staff was striking: "This isn't a factory preset. This is the result of living together."

"Give an agent a door, and it remains a guest"

For the past two years, every phone maker has been chanting "AI phone." But look closely and it's the same old story: a traditional OS with an AI shell on top—summarizing notifications, retouching photos, drafting replies. The paradigm hasn't changed. Humans operate phones, phones execute commands.

STEPX Neo takes a different path. StepFun built Step AOS, an agent-native operating system, from the ground up. As CEO Yin Qi put it: "Give an agent a door in the old system, and it remains a guest. Build it a house, and it becomes a resident."

The system has three architectural layers. At the compute layer, it dynamically orchestrates CPU, GPU, and NPU resources across on-device and cloud models. Lightweight tasks like setting alarms or finding photos stay on-device, completing in under 100ms with over 99% success rate. Complex reasoning and multi-step planning are handled in the cloud—while private data never leaves the device. At the data layer, perceptual, behavioral, and personal data are synthesized into unified semantic files served by a high-speed hybrid database. At the application layer, an atomic capability engine deconstructs system functions into composable services the agent can orchestrate. This means Amoo doesn't just "open apps"—it completes cross-app tasks the way a human would.

The triad: Memory, Execution, and Trust

If there's one sentence to distinguish STEPX Neo from conventional AI phones: traditional AI phones help you find answers; an agent phone helps you get things done.

Three capabilities make this possible.

Memory. Step AOS implements a dual-domain, three-step memory architecture. The user domain stores preferences and relationships; the agent domain accumulates experiential knowledge. The memory pipeline follows a "record—organize—recall" pattern: capture during interaction, consolidate in the background, retrieve precisely when needed. It achieves SOTA on benchmarks like Personamem and LongMemEval, with recall latency as low as 15 milliseconds for everyday queries. You never have to tell it "I don't eat seafood" twice.

Execution. The brain behind all this is StepFun's "1+N" model matrix. The newly released Step Edge on-device foundation model ranks first among its class across 29 authoritative benchmarks. Its end-to-end task success rate places it in the top tier of the AndroidDaily benchmark, and it holds the top domestic spot on Claw Eval.

Trust. Letting an agent spend your money and make decisions on your behalf demands trust. Step AOS implements a "trustworthy, visible, controllable, reversible" security framework: operations run in trusted execution environments, every step is auditable and traceable, permissions are granted on demand and revoked when done, and mistakes can be undone with a single tap. STEPX Neo is the first—and currently only—device to receive L3 certification under China's new national standard for AI terminal intelligence grading. StepFun also co-published a security whitepaper with the Shanghai AI Laboratory and contributed to the national standard for agent system security.

"See you in 100 days"

No hardware specs, no pricing, no handset partnerships were announced at launch. But the ecosystem is already taking shape: Ctrip, Alipay, Didi, Meituan, WPS, CapCut, Baidu, Amap, and JD.com are among the first-wave partners.

Yin Qi made a notable promise: apps won't disappear—they'll evolve into agent-native forms. "In 100 days, we'll reveal the full agent ecosystem and user experience at the second-half launch."

Meanwhile, StepFun kicked off the "STEPX Vision Program" on Bilibili, giving 1,000 content creators early access to incubate the first wave of native agent applications. The program opens to the public in mid-August with a call for feature requests and community-built Agents and Skills.

Not replacement, but co-evolution

There's a telling moment from the booth demo. A staff member sent a five-person group chat screenshot to Amoo. The system parsed conflicting constraints—budget limits versus accommodation preferences—generated a solution with confirmation checkpoints, and ultimately completed bookings and calendar sync, all in one flow.

This isn't "help me search for hotels" on steroids. It's an agent that closes the loop.

During his keynote, Yin Qi said: "The future is not a world where machines replace humans, but one where humans and intelligence co-evolve."

When a phone starts remembering you, understanding you, and getting things done for you—it stops being just a phone.