StepFun held an event in Shanghai on July 13 and introduced four products in one shot: the AI terminal brand STEPX, the agent-native operating system Step AOS, the personal agent Amoo, and STEPX Neo. The WeChat account “Banmianzhiwai” described it as one of the most “perfect” AI launches of the year, while arguing that the level of completeness itself made the event feel unusual.
The timing sat at the center of that reading. WAIC, the World Artificial Intelligence Conference, was set to open four days later. On July 17, Nubia and ByteDance were also due to unveil the mass-production version of the second-generation Doubao phone at WAIC, with the stated goal of selling it openly. In that context, the article says StepFun was not simply releasing products. It was trying to move first in the contest to define the “world’s first large-model-native agent smartphone.”
More than a launch, the event was framed as a bid for naming power
According to the article, StepFun chairman Yin Qi said he was not competing, but the structure of the event pointed in another direction. Its central task, the author argues, was to seize definition power. In emerging categories, the first company to frame the product can shape the standard that follows.
The clearest signal, in the author’s view, came from three lines shown at the event: “If you start too late, there is no point in doing it. If you start too early, the effort may come to nothing. If you do nothing, there may be nothing left to do later.” The piece argues that StepFun was not racing for four days on a calendar. It was trying to plant the first flag in the AI phone era.
StepFun used the event to answer the doubts raised around Doubao phones
The article says the more important detail was how complete the presentation looked. For a startup making its first public push into terminals, StepFun did not center its story on camera specifications or chip benchmark scores. It talked instead about models, systems, hardware, agents, ecosystem design, partners, permission management, trusted execution environments, audit trails, one-click rollback, deletable memory, security white papers and national standards.
That structure was notable because, as the article puts it, the company appeared to answer in advance many of the questions the market had thrown at ByteDance around the first Doubao phone. Those concerns included blurry boundaries around agent permissions, a joint blocking of simulated operations by Meituan, WeChat and Taobao, unclear safeguards for user privacy, reluctance by apps to open their interfaces to agents, and uncertainty over regulatory attitudes.
StepFun’s responses were laid out point by point. A TEE, or trusted execution environment, addressed whether the product was safe enough to use. Auditable and traceable steps addressed whether users and partners could see what the agent was doing. Permissions granted on demand and withdrawn once a task was complete addressed control. A one-click undo function addressed operational mistakes.
The response was not limited to product architecture. StepFun and the Shanghai Artificial Intelligence Laboratory jointly released the Security Technology White Paper for Next-Generation Agent Systems and the Cybersecurity Guidelines for On-Device Large Models. The article says those documents were the first to set out an agent security framework in a systematic way, and adds that the two sides were working together on national standards.
The first group of ecosystem partners was also highlighted: Trip.com, Alipay, Didi, Meituan, Amap, JD.com, Baidu, Weibo, WPS and CapCut. From the author’s perspective, this suggested StepFun had already completed its first round of permission negotiations, persuading apps to join as formal skills instead of being treated as external tools that rely on forced simulated clicks.
That is why the article describes the event less as a consumer product launch than as a polished defense presented to the industry. The actual device is not on sale. There was no pricing, no hardware specification sheet, no sales date, and no hands-on image circulating online, according to the piece. The real audience, it argues, was the app ecosystem, regulators, the capital market and the rival due on stage four days later.
In the article’s view, Yin Qi is racing the clock more than ByteDance
The analysis says many people may see this as a direct showdown between StepFun and ByteDance in AI phones. It rejects that framing. The company’s real source of urgency, the author writes, is time.
To make that case, the piece traces the condition of China’s large-model startups over the past few years. In 2023, the race was about who could build China’s GPT-4. In 2024, the focus shifted to benchmark scores and multimodal performance. By 2025, the story had changed. DeepSeek, through open source, free access and low-priced APIs, had reset industry rules. The article says its performance was benchmarked against OpenAI while pricing was only 1/30 as high. At the same time, Alibaba invested more than 100 billion yuan in a year, and ByteDance’s capital expenditure reached 160 billion yuan. Lee Kai-fu was cited as saying China’s large-model market would eventually move toward oligopoly, with only three major players left.
The article then moves into 2026. It says the once-mentioned “AI Six Tigers” had already split into very different paths. Zhipu and MiniMax had listed in Hong Kong and were now facing stock-price pressure. Moonshot AI was raising about $1 billion in a new round at a valuation of $18 billion. Baichuan was digging deeper into healthcare verticals.
StepFun, by contrast, appears in the article as a company in transition. Citing public information, the piece says StepFun’s pre-IPO valuation rose from $4 billion to $6 billion, its latest financing reached $2.5 billion, it had dismantled its red-chip structure, and it was pushing toward a Hong Kong IPO.
But the article argues that the figures look less secure once the business is broken apart. Most of StepFun’s revenue, it says, comes from acting as an AI supplier to smartphone makers. The piece adds that StepFun models are installed in 60% of China’s leading handset brands, with a total installed base of more than 42 million devices.
That can work in the short run. Over a longer period, the article says, the business becomes fragile because handset makers can replace StepFun at any time. Once on-device model capabilities start to converge, an AI supplier risks turning into the equivalent of MediaTek in the chip sector: productive, but with little pricing power.
The pressure runs in parallel with other shifts. API prices are falling. Model performance is converging. Consumer users are reluctant to pay. Enterprise customers mostly want customization. Large platforms still hold the ecosystem, the entry points and the data flywheel. The article’s bottom line is blunt: the main risk for independent model companies is not that their models fail, but that models themselves become less valuable.
From model company to terminal platform
The article ties that view closely to Yin Qi’s own background. It notes that he spent more than a decade at Megvii, taking the company from startup stage through multiple listing attempts before eventually leaving. Megvii had technology, commercial deployment and government and enterprise customers, the article says, but struggled for a long time with cash flow and never fully established a consumer closed loop.
In the author’s framing, Yin once believed in algorithms. At StepFun, he now believes in entry points. The article says his first decision after arriving at the company was to give up project-based ToB work, avoid a pure cash-burning ToC route, and bet on AI plus terminals instead.
It quotes Yin’s remarks in a media group interview after the July 13 STEPX launch: for independent foundation-model startups, pure ToC and pure ToB routes both fail the commercial math, while using terminals to drive commercialization is a more sustainable path. The article says that is not merely a strategy choice. It is a survival choice.
Under that logic, STEPX was not really just launching a phone. It was presenting a new valuation framework. StepFun was trying to reposition itself from a model supplier into a terminal platform.
A 100-day countdown
The article says StepFun cannot afford to wait. If the second-generation Doubao phone reaches mass production first, the right to define the category of the first AI agent smartphone may no longer belong to StepFun.
And the field does not end with one competitor. The piece points to Apple’s continuing work on Apple Intelligence, Huawei HarmonyOS Next’s deep integration with the Pangu model, Xiaomi’s second-generation AI system, and terminal strategies from both Tencent and Alibaba. By that stage, the author argues, startup companies may no longer have room at the main traffic and scenario entry points.
Another quotation from Yin is used to support that point: “If we ourselves cannot take the lead in building such an innovative terminal right now, then it will be very hard for our Step AOS to form a value closed loop, and it will also be very hard for consumers to use it.”
The meaning, as the article interprets it, is straightforward. Without self-built hardware, the operating system has no carrier. Without a carrier, there is no user data. Without user data, the agent cannot learn through task execution.
That is why, in the article’s reading, StepFun has to define in advance what an Agent Phone is, what an Agent OS should look like, and who gets to write the safety standard for agents. Even if the product is not fully mature yet, and even if the device appeared only on stage rather than in a hands-on area, the company still needed to occupy the position now.
Yin also said after the event, “The second half will be 100 days later.” The article links that timeline to capital markets. Before filing in Hong Kong, it argues, StepFun needs to upgrade its story from a model supplier to an intelligent terminal platform. It also mentions what it calls the company’s “$10 billion valuation ambition,” adding that API revenue alone cannot support that narrative and that a broader ecosystem story is needed.
What the author sees across China’s AI startups: launching the future
The final section widens the lens. StepFun’s event, the article says, is one example of a larger pattern. Nearly all Chinese large-model startups are now “launching the future.”
Zhipu is telling a capital-markets story around becoming the first publicly listed large-model company. MiniMax is focused on global consumer growth. Moonshot AI keeps emphasizing that overseas revenue has surpassed domestic revenue. Baichuan is presenting the scale of the healthcare vertical opportunity. StepFun, in this framing, has launched a phone that cannot yet be bought, an operating system that has not fully run through its use case, and an agent ecosystem that still needs another 100 days of work.
The author’s argument is that “today” no longer carries much value on its own. API calls, benchmark scores and monthly active users can all be overwhelmed by larger companies. What still attracts capital, the piece says, is the platform, the ecosystem, the operating system and even the role of a standards setter.
That leads to what the article calls an “over-allocation” phenomenon. Product maturity may be only 30%, but the completeness of the launch can reach 120%. Real user experience that can be delivered today may be limited, yet the story itself is told almost without gaps.
For the author, that is the posture available to startups in the current window. Capital is no longer investing in a model alone. It is investing in the possibility of a future platform. In that sense, Yin is trying to pull StepFun out of the old “AI Six Tigers” grouping and place it inside a different narrative track.
The article closes with a broader observation. All large-model startups, it says, are now trading future blueprints for present survival space. No one is waiting for a product to become perfect before making a move, because by then, there may no longer be a seat left at the table for startups. That is why this unusually polished StepFun launch is presented not as proof of comfort, but as evidence of urgency.