AI is bound to revolutionize the smartphone industry, but it will not be Stepverse.
But a forward-thinking concept does not equal industry subversion. A marketing-packed launch event, a pre-production prototype fully manufactured by a third-party OEM, and an agent system unvalidated by the market are far from enough to underpin the grand narrative of "overthrowing the mobile phone industry".
At a launch event in Shanghai on July 13, Stepfun made a high-profile debut of its self-developed terminal brand STEPX, unveiling the STEPX Neo, which it claims is "the world's first LLM-native agent phone". Standing on stage, Qi Ying painted a grand blueprint to disrupt the smartphone industry: the self-developed Step AOS native AI operating system, Amoo — an agent capable of autonomous cross-app execution, and a trinity of model-software-hardware, vowing to restructure human-computer interaction logic and revolutionize traditional mobile phones.
Off the stage, the capital audience burst into thunderous applause, and industry self-media outlets flooded online with coverage overnight. Press releases claiming to "disrupt the iPhone and outperform Android vendors" spread across the entire internet. But peeling back the hype and packaging of this noisy launch event, we must raise a sharp and objective question: with nothing more than an OEM-produced phone, a market-unvalidated agent system, and a large model that is not top-tier in the industry, does Stepfun truly have the ability to overturn the mobile phone industry landscape?
The answer is straightforward: the AI wave will eventually subvert the smartphone era, but the protagonist of this revolutionary upheaval will never be Stepfun. It may be the first player to step forward and take a bite of the fruit, but it is destined not to have the last laugh.
01 The Threshold to Revolutionize the Mobile Phone Industry Extends Far Beyond Large Models
To understand the limitations of Stepfun, we must first look at the globally sensational lawsuit between Apple and OpenAI on the other side of the ocean.
Not long ago, Apple formally filed a lawsuit against OpenAI in federal court. The surface of the complaint accuses OpenAI of systematically poaching more than 400 software and hardware engineers and stealing trade secrets related to terminal hardware heat dissipation, on-device chips, and sensor arrays. But at its core, this is a survival war for the next-generation computing entry point.
Apple's fear is palpable: the moat of the smartphone ecosystem built by the iPhone is rooted in the closed-loop iOS software-hardware integration. Once a company that masters top-tier large models, independently develops terminal hardware, and can run autonomous agents without relying on iOS/Android emerges, Apple's decades-long dominance over the mobile internet entry point will completely collapse. This is the fundamental reason why Cook did not hesitate to break with past cooperative relations and strike hard against OpenAI.
Many people simply assume that "a large model alone can disrupt the mobile phone industry", but Apple has proven an unshakable industry rule through its actions: a player capable of revolutionizing the mobile phone industry must hold three aces at the same time: top-tier general large model capabilities, complete terminal software and hardware engineering experience, and mature and controllable full-domain ecosystem dominance. None of the three can be missing — any missing link will keep a company stuck at the "concept marketing" stage.
Looking back at Stepfun, it does not have a firm grip on any of the three aces.
First, the first ace: top-tier large model capabilities. The domestic large model track has long been stratified. In terms of comprehensive reasoning, coding, long-text processing, and end-user reputation, DeepSeek has built a massive developer ecosystem through its open-source approach, Kimi has firmly captured office users with its ultra-long context and document processing capabilities, while overseas Claude and GPT-4o are firmly in the global first tier.
Stepfun's Step3 series models only have local advantages in the niche scenario of lightweight reasoning on domestic chips. They have never reached the top tier in comprehensive benchmark tests or the actual experience of mass users. Its previous core business model was to export on-device AI middleware to mobile phone manufacturers such as Xiaomi and OV, acting as a supporting service provider dependent on hardware vendors, never having an independent massive end-user base, and its user mindshare is far less solid than that of its competitors.
The second ace: complete terminal software and hardware engineering experience — this is precisely Stepfun's biggest weakness. The mobile phone industry has never been an asset-light track where "algorithms are enough to make hardware". It is a capital-intensive, long-chain, low-fault-tolerance heavy industry. For a single mobile phone, every link from structural design, screen calibration, heat dissipation stacking, battery packaging, supply chain management, quality control testing, to offline channels and after-sales maintenance requires years or even decades of industry accumulation.
Stepfun's entire core team comes from the fields of algorithms and computer vision. Since its founding, the company has never been involved in hardware manufacturing. The newly released STEPX Neo is fully ODM-manufactured by Huaqin Technology, with the complete machine design, supply chain, and mass production quality control all outsourced. Stepfun is only responsible for adapting the AI system and model, which is essentially OEM customization rather than independent complete machine R&D.
In contrast, manufacturers such as Huawei and Xiaomi, which have deep roots in the terminal industry for more than a decade, hold self-developed chips, tens of millions of levels of supply chain bargaining power, offline stores across the country, and a complete after-sales system. Even OpenAI, which crossed borders to enter the hardware field, spent heavily to poach Apple's core hardware team, acquired Jony Ive's design studio, and built its own hardware R&D team from scratch, rather than simply finding an OEM for branded production. Starting from zero, Stepfun skipped all hardware accumulation links and claimed to disrupt the industry with nothing more than a prototype made by an OEM — that is far too idealistic.
The third ace: full-domain ecosystem dominance. The core competitiveness of an agent phone lies in its ability to autonomously perform complex tasks across apps. To achieve "booking a ticket with one sentence, notifying colleagues synchronously, and automatically generating documents", the prerequisite is that major internet applications are willing to open underlying atomic permissions to the system.
At this launch event, Stepfun announced a number of ecological partners including Meituan, Alipay, and WPS, but the cooperation is only at the shallow signing stage, with no underlying system-level integration achieved. The biggest practical obstacle in the current industry is that major internet platforms will never easily grant core operation permissions to third-party terminal systems — once Stepfun's Amoo agent can automatically operate Taobao, WeChat, and Douyin, it is equivalent to platforms giving away their user traffic and transaction initiative, touching the core interests of internet companies.
Huawei's HarmonyOS and Xiaomi's HyperOS can achieve cross-device collaboration relying on their massive terminal installed base, forcing platforms to adapt for the sake of their huge user base. But Stepfun, with a new niche phone whose price is not even announced and is not officially on sale, has no user scale as a bargaining chip for negotiations. Its ecosystem cooperation can only stay on the launch event PPT, and after the product is launched, it is very likely to face the awkward situation where "the agent understands the instruction, but the app refuses to grant permissions to execute".
Having obvious shortcomings in all three core aces means that the path for Stepfun to replicate OpenAI's disruption of Apple is fundamentally unfeasible.
02 Native Agent System: A Paper Blueprint or a Real Implementation?
The most eye-catching concept at the entire launch event is "Step AOS, the world's first LLM-native operating system". The officials repeatedly emphasized that traditional mobile phones are "Android systems with external AI tools", while Step AOS reconstructs resource scheduling logic for agents at the bottom layer, realizing full-link autonomous operations of perception, memory, decision-making, and execution, and completing the paradigm shift from "people operating apps" to "agents autonomously handling tasks".
This narrative sounds flawless, but as long as we consider the common pain points in the current industry-wide implementation of agents, we will find insurmountable engineering gaps hidden behind it.
First, the industry ceiling for agent task completion rates. According to the latest industry benchmark test data, even the world's most cutting-edge LLM agents only have a 37% comprehensive completion rate when handling multi-step complex tasks across apps. Nearly 60% of ten commands will encounter problems such as logical breaks, operation errors, and permission failures. This stability problem that even top model manufacturers with years of experience cannot solve is difficult to prove as stably implementable on a large scale by Stepfun with nothing more than a demo video at a launch event.
The demo scenarios at the launch site were all rehearsed in advance in an ideal, interference-free environment. In real user usage, any variable such as app version updates, system permission restrictions, network fluctuations, and account security verification will directly invalidate Amoo's fully automatic operations. More critically, the compliance boundaries for AI agents to automatically operate third-party applications are not yet clear in China. Fully automatic simulated clicks and cross-platform access to user data carry regulatory risks: aggressive operations can easily trigger platform risk control or even compliance penalties, while conservative operations will greatly weaken the core experience of agents. This industry dilemma with no way out is unavoidable even for Stepfun.
Second, the "impossible triangle" contradiction between hardware computing power, power consumption, and heat dissipation. A native agent requires a large model to continuously run in the local background of the phone, placing extremely high demands on the phone's memory, computing power, heat dissipation, and battery life. The hardware architecture of traditional Android phones is designed for video, social media, and games, not optimized for 7×24 hours of local large model reasoning.
Stepfun Neo only adds a small interactive secondary screen on the back of the device, with no publicized self-developed chips, customized heat dissipation structures, or high-capacity battery solutions. The complete hardware architecture still follows the general template of mainstream mid-range mobile phones. When the Amoo agent runs in the background for a long time, serious overheating, a sharp drop in battery life, and system lag will inevitably occur. The drawback that model manufacturers lack hardware calibration experience will be fully exposed during long-term user usage.
In contrast, Huawei and Xiaomi's AI mobile phone routes adopt a prudent path of "terminal hardware first, with AI capabilities gradually adapted": after polishing their hardware supply chains for more than a decade, they develop self-developed on-device chips to optimize AI computing power and power consumption, then iterate large models and agent functions, with the hardware bottom layer serving AI. But Stepfun completely reverses this order: it first builds a concept of an AI-native system out of thin air, then finds an OEM to match existing hardware. This upside-down R&D order is destined to prevent the hardware from fully unleashing the model's potential.
In addition, there is a core issue easily overlooked: the lack of a closed-loop business model. The mobile phone industry is a typical scale-driven industry with thin hardware profits, which must rely on shipments of tens of millions of units to amortize R&D and supply chain costs. Stepfun's previous main business was to export AI model services to B-end manufacturers, with no experience in consumer electronics retail channels, offline brand stores, or user operation teams. It has neither a mature distribution network in the mobile phone industry, nor end-user brand operation experience.
Even if STEPX Neo is successfully launched, it can only be sold through niche online channels, and its shipments are unlikely to exceed the million-unit level. Meager sales cannot amortize the costs of complete machine R&D and system iteration. Long-term losses in the hardware business will continuously consume the company's cash flow, in turn squeezing investment in underlying large model R&D, falling into a vicious cycle where "loss-making hardware drags down algorithm development, and insufficient algorithms lead to poor hardware sales".
03 Who Is the Real Player Capable of Revolutionizing the Mobile Phone Industry?
We do not deny that AI will eventually disrupt smartphones, but participants in this transformation have long been divided into two clear core tracks — and Stepfun has no seat in either route.
The first route: Large hardware manufacturers develop large models in-house, achieving deep two-way software-hardware integration, with representative players Huawei and Xiaomi.
Such manufacturers hold full-chain experience in terminal hardware, with self-developed chips, massive terminal installed bases, mature offline channels, and strong ecological negotiation chips. They follow the steady path of "hardware as the foundation, AI as the soul". With more than a decade of deep involvement in the mobile phone industry, they are well aware of the real pain points of end users, can optimize the computing power and power consumption of on-device large models at the chip bottom layer, and achieve deep system-level adaptation with WeChat, Alipay, and major content platforms. Their agent functions iterate steadily, with no fatal weaknesses such as OEM production, disconnected ecosystems, or missing channels.
Huawei's HarmonyOS with the Pangu large model, and Xiaomi's HyperOS with the Xiaoai agent, are both polished and optimized based on hundreds of millions of existing mobile phone users. Their AI functions fit daily usage scenarios of the public, with extremely low implementation risks. The advantage of such players is that they do not need to create a brand new mobile phone category out of thin air — they only need to continuously superimpose AI agent capabilities on their existing huge hardware base, gradually completing the transition from smartphones to AI-native terminals, with low transformation costs, solid user bases, and extremely strong risk resistance.
The second route: Top-tier general large model giants make heavy full-stack investments in self-owned hardware, with representative players OpenAI and DeepSeek.
Such players hold top-tier large model technology in the industry, with massive end-users and developer ecosystems, and sufficient financing cash flow to support long-term hardware R&D. Take OpenAI as an example: it spared no expense to poach hundreds of Apple hardware engineers, acquired top industrial design studios, and planned self-developed on-device AI chips, building a complete hardware R&D team from scratch. It is by no means simply looking for an OEM for branded production, but is willing to invest years of time to accumulate terminal hardware capabilities and fill the gap in software-hardware integration.
Domestically, DeepSeek has accumulated the world's largest developer community with its open-source models, and its end-user reputation and comprehensive model capabilities are firmly in the domestic first tier. Once it launches hardware layout, it has sufficient technical confidence and user base to support terminal product R&D. The core competitiveness of such players lies in their irreplaceable top-tier model technology, with sufficient time and capital to fill hardware gaps, possessing the underlying technical strength to disrupt the traditional mobile phone industry.
In contrast, Stepfun is neither a hardware giant in the first route, nor has it reached the technical height of the top-tier large model giants in the second route, stuck in an awkward middle zone: zero hardware accumulation, model capabilities unable to enter the industry's first tier, and obvious shortcomings in the full chain of ecosystem, channels, and after-sales services, making it impossible to gain a firm foothold in either transformation track.
It is more like an "experimenting pioneer" in the industry, taking the lead in exploring the path for large model manufacturers to make mobile phones, clearing the track risks for real giants such as Huawei, Xiaomi, and OpenAI. But a pioneer is not equal to a victor — the first player to enter the market is often a stepping stone for later giants.
04 Conclusion
Behind the noisy launch event, do not confuse "concept advancement" with "industry subversion".
It is undeniable that Stepfun's launch event has indeed provided new imagination for the AI terminal industry. Its courage to break away from the traditional mobile phone idea of adding AI features, and propose a new concept of LLM-native agent terminals, is worthy of recognition.
But a forward-thinking concept does not equal industry subversion. A marketing-packed launch event, a pre-production prototype fully manufactured by a third-party OEM, and an agent system unvalidated by the market are far from enough to underpin the grand narrative of "overthrowing the mobile phone industry".
Apple fears OpenAI because the latter holds the world's most powerful large model, and is willing to spend heavily to fill the full chain of hardware capabilities. Huawei and Xiaomi steadily iterate AI mobile phones relying on more than a decade of terminal industry accumulation, promoting progress through software-hardware collaboration. But Stepfun has neither a top-tier model powerful enough to disrupt the industry, nor years of deep-rooted hardware industry accumulation. Every link from ecosystem, channels, after-sales service, to quality control has irreparable shortcomings.
AI will definitely rewrite the future of smartphones, and the next-generation AI-native terminals will surely redefine human-computer interaction. But the protagonist who completes this industrial revolution will definitely not be Stepfun. It can be the first to taste the fruit, but it is difficult to become the ultimate winner.
This article is from the WeChat public account "Coopetition AI", author: Coopetition, published with authorization from 36Kr.