Chinese foundation-model developer StepFun has done what most model companies only talk about: it launched its own phone. At a "Ready Builder One" event in Shanghai on October 13, the company's terminal arm STEPX unveiled the STEPX Neo, a smartphone designed from the ground up around a personal AI agent rather than apps — one of the first big-model-native phones from a Chinese model vendor going direct to hardware.
The hardware is built around sensing and expressing. The Neo carries a six-microphone array, a Whisper 1.0 pickup algorithm and a dedicated low-power audio sensing chip to catch the user's voice in noisy rooms, multi-speaker conversations or whispered commands. A dedicated AI key on the side wakes the assistant, and an independent MiniLED dot-matrix display on the back shows what the agent is doing — waking, listening, thinking. The design language is described as "retro-future," with five easter eggs honoring milestones in AI history, in carbon black or silicon silver.
The software is the real story. Step AOS 1.0 is billed as the first agent-native operating system built from scratch, with long-term memory, task scheduling and permission management built into the system layer. Memory is split into two domains — a user domain recording personal information, experiences and preferences, and an agent domain accumulating knowledge, methods and know-how. A "trustworthy, visible, controllable, reversible" architecture makes task processes auditable and lets users manage permissions and memory, with the company stressing that final control stays with the user.
Execution is split between device and cloud: on-device models handle quick local commands such as adjusting settings or setting alarms, some of which work offline, while cloud models plan and decompose complex multi-app tasks and hand them back to the device to execute. Those capabilities are company claims — independent reviews of the phone have yet to surface.
For developers, STEPX introduced Step Studio, its first AI-terminal-native IDE: creators describe what they want in natural language and get AI apps, games, agents and Skills, with access to the phone's sensors, camera and rear display, plus MCP and A2A protocols to connect external services and specialist agents. At the event, Bilibili creators showed a flight-monitoring and decision-support system for a homemade flying skateboard, and a "mind mouthpiece" Skill that connects a brain-computer interface to the phone to help people with speech difficulties express emotions.
The ecosystem play is where the phone gets interesting. StepX says the Amoo assistant has tied up with nearly 100 partners, including Alipay — Amoo can invoke Alipay's services after user authorization, and STEPX is the first terminal vendor to integrate Alipay's Vibe Pay, letting developers build transactions into apps made in Step Studio. Tencent products in the first wave include WorkBuddy, WeCom and WeChat Read; CapCut (Jianying) collaboration brings A2A cross-agent video editing; and Ctrip, Didi, Amap and Meituan round out local services.
Pricing is aggressive by any standard. The launch edition costs 3,999 yuan (about US$560) as a full "software-hardware-model" stack, bundling 12 months of token quotas at 10 billion tokens per month on top of the device. The first 10,000 buyers who complete qualifying tasks get a "Starship Ticket" entitling them to a zero-cost trade-in for a future STEPX device within five years. Pre-orders opened at 9 p.m. on October 13 across JD, Tmall and Douyin, with official sales starting October 16.
The launch marks the next phase of the agent-platform race: model vendors moving downstream to own the entry point where agents meet users. Bundling tokens into the hardware price mirrors how US assistants monetize through subscriptions, and the partner integrations — payments, mobility, office, video — show StepFun betting that an agent is only useful if it can actually complete transactions. Whether agent-native phones move beyond novelty will depend on execution and developer uptake, but StepFun has at least put a concrete price and a date on the experiment.
Comments (0)
Log in to join the discussion
Log InNo comments yet