This might be the most "perfect" product launch in the AI industry this year.
So perfect that it feels a bit unusual.
On July 13, Step Fun held a launch event in Shanghai, unveiling four new offerings at once: the AI terminal brand STEPX, the intelligent-agent-native operating system Step AOS, the personal intelligent agent Amoo, and the world's first large-model-native intelligent agent phone, STEPX Neo.
Four days later is the WAIC World Artificial Intelligence Conference.
And on July 17, Nubia and ByteDance are set to launch the mass-produced version of the second-generation Doubao Phone at WAIC, this time with the clear intention of selling it widely.
Timing is definitely not a coincidence.
While Step Fun's Chairman Yin Qi said there's "no competition," the entire launch event was about one thing: claiming the right to define.
There is only one crown for being the world's first large-model-native intelligent agent phone. Whoever claims it first gets the chance to hold the industry standard in their hands.
Those three lines shown at the event were more candid than any product specification.
"Do it too late, and you won't get to do it at all. Do it too early, and you might do it for nothing. Don't do it, and soon there will be nothing left to do."
What Step Fun is seizing is not just four days, but the first flag of the AI phone era.
I. Blocking Every Pitfall Doubao Stepped Into
What's more noteworthy is the completeness of this launch.
A startup, making its first public appearance with a terminal product, didn't compete on camera megapixels or chip benchmark scores. Instead, it discussed model, system, hardware, intelligent agents, ecosystem, partners, permission management, Trusted Execution Environment (TEE), operation auditing, one-click undo, deletable memory, security white paper, national standards.
If you followed the industry discussions around the Doubao Phone half a year ago, you'd find that Step Fun has preemptively addressed almost every criticism the market once directed at ByteDance.
Recall the list of issues exposed by the first-generation Doubao Phone.
Unclear permission boundaries for intelligent agents, leading to joint blocking of simulated operations by Meituan, WeChat, and Taobao. Lack of clear user data privacy guarantees. Apps unwilling to open entry points to intelligent agents. Unclear regulatory stance.
Now look at STEPX's approach.
TEE (Trusted Execution Environment) solves the 'dare to use' problem. Every operation being auditable and traceable solves the 'can you see it' problem. Permissions granted on-demand and revoked after use solve the 'can you control it' problem. One-click undo for misoperations solves the 'what if it's wrong' problem.
This isn't just a response on a technical level. Step Fun, in collaboration with Shanghai AI Laboratory, released the New Generation Intelligent Agent System Security Technology White Paper and the End-side Large Model Network Security Guide, systematically proposing an intelligent agent security framework for the first time and jointly developing national standards.
The list of first-phase ecosystem partners is also substantial. Ctrip, Alipay, Didi, Meituan, Gaode, JD.com, Baidu, Weibo, WPS, Jianying.
This means Step Fun has completed the first round of permission negotiations, getting Apps to proactively join in the form of 'skills,' unlike the first-gen Doubao, which was seen as an invasive "plug-in" forcefully simulating clicks.
This no longer seems like a product launch; it's more like a perfect thesis defense presented to the industry.
The audience for this defense isn't consumers. The actual phone isn't on the market yet—no price, no configurations, no release date, not a single hands-on photo to be found online.
The audience is the App ecosystem, regulatory bodies, the capital market, and the formidable rival set to appear four days later.
A startup, with its product not yet mass-produced, has preemptively addressed all future concerns and potential criticisms.
Why?
II. What Yin Qi Fears Most Is Not Losing to Doubao
Many might think this is a two-horse race in the AI phone arena. Step Fun versus ByteDance, a clash of two approaches.
That's not the case.
What Step Fun is truly anxious about is time.
To understand this, one must first see the real situation of large model startups today.
In 2023, the fight was about who could create China's GPT-4. In 2024, it was about who had the highest benchmark scores, who had the strongest multimodal capabilities.
2025, the story changed. DeepSeek reshaped industry rules with open-source, free access, and low-cost APIs, matching OpenAI's performance but priced at 1/30. Alibaba invested over 100 billion yuan annually, ByteDance's capital expenditure reached 160 billion yuan. Kai-Fu Lee said the Chinese large model market will ultimately move towards oligopoly, with only three players left in the end.
By 2026, the former "AI Six Tigers" have completely diverged. Zhipu AI and MiniMax have listed on the Hong Kong Stock Exchange, now facing stock price pressures. Moonshot AI is raising a new round of about $1 billion at a $18 billion valuation. Baichuan AI is focusing deeply on verticals like healthcare.
And Step Fun?
According to public information, Step Fun's Pre-IPO round valuation was pushed from $4 billion all the way up to $6 billion. The latest financing round raised a huge $2.5 billion. It has dismantled its VIE structure and is sprinting towards a Hong Kong IPO.
The numbers look decent. But upon closer inspection, problems emerge.
Step Fun's revenue mainly comes from being an AI supplier for smartphone manufacturers. Data shows that 60% of China's top mobile phone brands are equipped with Step Fun's model, with installations exceeding 42 million units.
Short-term, this business model works. Long-term, it's hanging by a thread.
Major Chinese manufacturers like Huawei, Xiaomi, OPPO, and Vivo could replace Step Fun at any time.
As on-device model capabilities become commoditized—and this is indeed accelerating—AI suppliers risk becoming the MediaTek of the chip industry. They have output, but zero pricing power.
Simultaneously, API prices are falling lower, model capabilities are converging, C-end users are reluctant to pay, and B-end clients all demand customization. The tech giants have ecosystems, entry points, and data flywheels.
The biggest risk facing large model startups today is not failing to build a good model; it's that the model itself is becoming less and less valuable.
Yin Qi understands this better than anyone. He spent over a decade at Megvii, from startup through multiple attempts to go public, until his eventual departure. Megvii had technology, had deployments, had government and enterprise clients, but its cash flow was perpetually under pressure, never closing the loop for a consumer-grade business.
A decade ago, at Megvii, he believed in algorithms. Today, at Step Fun, he's starting to believe in the entry point. His first major decision after joining Step Fun was to abandon pure ToB project-based work and pure ToC cash-burning, betting on AI + terminals. He was very direct in the media group interview after the July 13 STEPX terminal product launch: For independent base model startups, pursuing pure ToC or ToB paths doesn't make financial sense; using terminals to drive commercialization is a more sustainable path. This isn't a strategic choice; it's a survival choice.
What STEPX is truly launching is not a phone; it's a new valuation logic.
From a model company to a terminal platform.
III. The 100-Day Countdown
Today's Step Fun cannot wait, and cannot afford to wait.
Once the second-generation Doubao Phone enters mass production, the right to define the world's first AI intelligent agent phone is no longer in your hands.
Looking further ahead. Apple's Apple Intelligence continues to evolve. Huawei's HarmonyOS Next deeply integrates the Pangu large model. Xiaomi's AI system is entering its second generation. Tencent and Alibaba each have their own terminal strategies.
By then, there will be no place for startups at the entry points for traffic and scenarios.
Yin Qi said in an interview: If we ourselves cannot now take the lead in creating such an innovative terminal, it will actually be very difficult for our Step AOS to form a value loop, and also difficult for consumers to use.
The meaning is clear. Without making their own hardware, the operating system has no carrier. Without a carrier, there is no user data. Without user data, the intelligent agent cannot learn to perform tasks.
Therefore, today's Step Fun must pre-define what an Agent Phone is, what an Agent OS should look like, and who writes the security standards for Agents.
Even if the product isn't fully mature today, even if the real device only appeared on stage and not in a hands-on area, this stake must be claimed.
Yin Qi said one thing in the post-event interview: The second half will be in 100 days.
From a capital market logic perspective, before submitting its Hong Kong listing application, he must upgrade the valuation narrative from that of a model supplier to an intelligent terminal platform.
A $10 billion valuation ambition simply cannot be supported by API sales revenue; it requires selling a more grandiose ecosystem story.
Plant the flag first, even if the territory isn't fully secured.
IV. Chinese AI, Beginning to Oversell
Step Fun's meticulously designed, perfect launch is just one sample of the times.
What's truly worth observing is a more widespread underlying phenomenon.
Today, almost all Chinese large model startups are doing the same thing: launching the future.
Zhipu AI tells the capital story of being the world's first publicly listed large model company. MiniMax focuses on the growth story of global C-end expansion. Moonshot AI repeatedly emphasizes the miracle of overseas revenue surpassing domestic. Baichuan AI launches the imagination space in verticals like healthcare.
Step Fun is launching a phone that can't be bought yet, an operating system that hasn't proven itself, and an intelligent agent ecosystem that still needs 100 days of polishing.
"Today" has already depreciated. Today's API call volumes, model benchmark scores, monthly active users—all could be crushed by the giants at any moment.
The only thing that can still make the capital market buy in is platform, ecosystem, operating system, even being a standard-setter.
This inevitably leads to an interesting "over-hyping" phenomenon. The product maturity might be only 30%, but the launch's completeness is 120%. The real experience deliverable today is very limited, but the story is told almost flawlessly.
This is the only posture startups like Step Fun can choose in the current window period.
Capital is no longer investing in a model; capital is investing in a future platform.
For this story,
Yin Qi is pulling Step Fun out of the ranks of the "AI Six Tigers" and squeezing it into a different track narrative.
【Beyond the Layout】:
Today, all large model startups are doing the same thing. Trading today's survival space for a blueprint of the future.
No one is waiting for the product to be perfect before making a move. By that day, there might be no seats left at the table for startups.
Yin Qi said, do it too late, and you won't get to do it at all.
He certainly did this early. Holding a launch before the product is mass-produced, defining national standards while the OS still needs 100 days of polishing, declaring "world's first" while the phone is still being redesigned.
Step Fun's overly perfect launch precisely proves one thing.
In this brutal elimination race, the more flawlessly you articulate the future, the more it reveals how urgent the present situation is.
This article is from the WeChat public account "Beyond the Layout", author: Huahua
