StepFun shipped its own agent native OS, Step AOS, and the STEPX Neo phone in Shanghai on July 13, with Ctrip, Alipay, DiDi, and Meituan signed up as launch partners.
The product StepFun actually shipped on July 13 is not the STEPX Neo phone. It is Step AOS, an operating system whose primitive surface is agent-callable: every piece of hardware and every legacy app is exposed to the agent as compute, data, or a skill. The phone is the demo. The OS is the bet.
StepFun (雷峰网 launch coverage) unveiled the new STEPX brand, Step AOS, and the STEPX Neo phone at a Shanghai event on July 13. The headline framing, "the world's first large-model-native AI terminal brand," is the company's own category term and rests on a single Chinese-language launch report; English-language wire confirmation is still pending. The structural move, the OS, is verifiable and worth taking seriously on its own.
Step AOS is built around three stated pillars: memory, decision and execution, and security. The user states an intent; the agent handles the full chain. Underneath sits StepFun's 1+N model matrix, a single foundation model plus N multimodal specialists that give the agent sight, hearing, and physical-world understanding (Sina Tech). For terminal deployment, StepFun built Step Edge, an on-device model that runs light agent tasks locally with full-modality local privacy handling, and escalates complex tasks to the cloud flagship. The design principle the company names, "serves humans, also AI-friendly" (为人服务、也对AI友好), encodes the same shift: the terminal is no longer a passive responder but a partner that senses and acts.
The two design choices that actually matter are the memory architecture and the end-cloud split. StepFun claims a dual-domain memory design reaches SOTA-grade performance in an unspecified "international authority evaluation," and that daily Q&A memory recall runs in as little as 15 milliseconds. Both numbers are company-stated, not independently measured, and the "international authority" is not named; treat the figures as a launch claim, not a benchmark result. The end-cloud coordination, simple tasks answered locally, complex tasks escalated to the cloud, is the more durable piece: it is what lets a phone-sized device carry a full agent without melting its battery or its thermals.
The other signal worth tracking is the L3 pass. STEPX Neo cleared the first batch of China's new AI Terminal Intelligence Grading standard at L3, the highest level currently open, and is the only agent phone to receive the certificate (Sina IT). L3 is a self-submitted test against a national standard, not an independent third-party review, so the pass is real but its weight is the certification gate, not a quality endorsement. It does, however, give StepFun a concrete answer to "is this thing certified" that ByteDance's forthcoming Doubao phone will have to match or beat.
On the security side, StepFun and Shanghai AI Lab jointly released a New-Generation Agent System Security Technical White Paper and an On-Device LLM Cybersecurity Guide, proposing a "Trusted, Visible, Controllable, Reversible" (可信、可见、可控、可逆) framework and co-developing national agent-security technical standards. The framework is a policy move as much as a product move: it pre-positions StepFun's vocabulary inside the standard-setting conversation before ByteDance or any other phone-side agent maker gets a seat at the table.
The partner list is the proof of work. Ctrip, Alipay, DiDi, Meituan, WPS, and CapCut are signed up as launch partners, covering travel, government services, local life, office productivity, and content creation (anzhuo.cn coverage). An agent that can call Ctrip to rebook a flight, Alipay to settle a bill, and DiDi to send a car, in one intent, is a different product from a chatbot with an API key. The launch program Bilibili is hosting, the STEPX Fantasy Plan, opens user wish-gathering and agent/skill co-creation to the public in mid-August, which is the first dated next milestone worth watching.
StepFun is not making this bet in a vacuum. Sina framed the launch as "rushing ahead of the Doubao phone," placing it in direct competition with ByteDance's own AI phone program. The structural question is identical: when the OS primitive is an agent and not an app icon, who owns the surface area, the AI lab, the phone maker, or the OS gatekeeper? StepFun is choosing the AI-lab-first path, shipping both the model and the OS, with Ctrip, Alipay, DiDi, Meituan, WPS, and CapCut signed up as launch partners. ByteDance's answer is forthcoming. The watch item is not which phone has a better camera; it is which agent handles a Ctrip-to-Alipay handoff in fewer taps, and which one does it without the user opening an app at all.