By
Edited by NUPIAO Editorial Team
In the first half of the year, as leading companies went public and secured funding one after another, the heat in the AI large model space kept rising. Among the many players, Jieyue, which had been relatively low-key, is trying to grab more attention with a fresh new path.
On July 13, Jieyue unveiled its Agentic-native OS, Step AOS, and simultaneously launched its personal AI agent, Jieyue Amoo, the large-model-native AI terminal brand STEPX, and the large-model-native agent smartphone STEPX Neo. This completes a full-stack layout from large models and operating systems to terminal hardware.
Why does a model company feel the need to dive into making terminal hardware? Actually, Jieyue had already been in the smart terminal business for cars and phones, playing its part in the industry chain as a model supplier.
Yin Qi, Chairman of Jieyue, told NUPIAO that it’s because, in the current industry ecosystem, it’s tough for the company to present its large models, operating system, and AI agent to users in a complete form. So, they decided to first land a full, end-to-end experience on a smartphone, as a pioneering exploration.
“At the same time, we have a very open attitude. In the future, we’re willing to open up our agent and operating system to other hardware OEMs,” Yin Qi said.
As the competition in large models gradually shifts from model capability to agent deployment, AI operating systems are poised to become the core infrastructure for the next phase of the AI race.
Jieyue believes that current mobile OSes are built on the logic of human-operated apps, which makes it hard to support core agent abilities like long-term memory, autonomous decision-making, and cross-app execution. That’s why they proposed rebuilding the runtime environment around the agent, rather than just layering AI features on top of a traditional OS.
Breaking down the permission barriers between apps on an agent phone is a challenge the whole industry faces. “Right now, we believe the most critical part is at the OS level—solving it with software,” Yin Qi said. He explained that the team has modified the Android Kernel for AI and plans to rewrite open-source ecosystems like Linux and RTOS in the future.
During a media interview after the launch, Yin Qi mentioned that building an AI terminal isn’t about entering the traditional phone market. It’s about creating a closed technical loop for the agent era. “If we don’t take the lead in making an innovative terminal, it’s really hard for our AOS to form a value loop, and it won’t be accessible to regular consumers,” he said.
Yin Qi pointed out that while today’s agents can handle tasks like coding and writing, their ability to perform complex operations on real devices like phones and PCs is still limited. The root cause isn’t a lack of model power, but the absence of a platform built for the agent to run on.
But he stressed that the company won’t become a traditional phone maker, nor will it make hardware sales its core metric. Although the business model isn’t crystal clear yet, he believes the future business model for AI terminals will be different from the current one. “It definitely won’t just be about selling hardware for a profit, and it certainly won’t be about making money from long-tail app pre-installation ads.”
Yin Qi acknowledged that “Doubao’s phone” is a really important exploration for the whole industry, bringing the concept of the agent phone into everyone’s line of sight.
However, he thinks that past definitions of Agent OS were more like iterative upgrades on top of Android. Jieyue’s vision for an OS is different: it builds a runtime environment for agents based on the principle of human-machine symbiosis, running compatibility layers under Android, Linux, and RTOS. “We believe that an operating system like this should be cross-device, not trapped on just one phone or one PC.”
“In terms of OS definition, we’re very different from existing terminal makers and Doubao,” Yin Qi said. “There’s no standard answer right now. The ideal state is for everyone to try things out quickly and see which path is more correct.”
Beyond this, Yin Qi also answered many questions about Jieyue’s roadmap for the next phase, which is the biggest thing people are curious about since he took the chairman role.
The following is a lightly edited transcript of the interview with Yin Qi:
Q: Jieyue previously deployed its tech in car and phone hardware by providing on-device large model capabilities to OEMs. What’s the most fundamental difference between launching an AI terminal brand now and that earlier strategy? Why does a model company have to go down this road?
Yin Qi: That’s a great question. Before, we were in a ToB model, providing a full suite of AI capabilities to phone and car OEMs. We’ll continue that. But in the agent era, and within the current phone ecosystem, it’s hard for us to present our model, OS, and the whole agent product in a complete way on an existing phone.
We want to be a first mover, to build a complete system first. At the same time, we hold a very open attitude. In the future, we’re very willing to open up our agent and operating system to other hardware OEMs.
Q: Why did you choose a phone form factor for your first terminal? What’s the biggest challenge you’ve hit with the hardware?
Yin Qi: For the first question, I want to say that the Jieyue team has over a decade of experience in the hardware supply chain. We have a lot of respect for making hardware and we’re very cautious about it. Choosing to go for a closed loop like this ultimately came down to choosing to do the hard but right thing. Today’s launch had a lot of content, but the main event was our Step AOS. Our first step was to build the OS. If we don’t take the lead in making such an innovative terminal, our Step AOS would have a hard time forming a value loop, and it would be tough for consumers to use. It’s a simple logic.
We talked about the future ecosystem. It’s really going to be symbiotic and open. I think if this Step AOS and its ecosystem can grow, we’d be very happy to empower existing phone makers with this Step AOS capability. This is a process of forming an industry chain.
Second point, the biggest challenge? Making a large model is hard, making hardware is hard. In a way, it’s ‘hard plus hard.’ What’s even harder than that is if we don’t even believe in the alternative paths for commercializing it. I think Chinese large model companies need to carve out their own path to commercialization and industrialization.
Q: You announced your formal joining of Jieyue early this year. It’s been almost two hundred days now. What strategic adjustments have you driven in that time? What resistance did you encounter, and how did you overcome it?
Yin Qi: Let me give you another example. The hottest company globally right now is Anthropic. But they set their coding strategy back in 2021. What we see today is the result of five years of work. For an LLM startup, you can’t just catch up later. You have to make choices, trade-offs, and have faith from day one.
Jieyue, from Day 1, believed in the combination of software and hardware. It believed that large models need a good hardware carrier to reach consumers. It’s a ToC story, not a ToB one, and that’s how we’ll form our future commercialization loop. That strategy was there from day one.
After I joined, the essence was to accelerate the execution of that strategy. That’s why Jieyue didn’t say much for the past three years. Now, we feel we’re ready to present our full strategic picture. Of course, it’s still going to be incredibly hard going forward, but that’s probably the truly interesting part of AI entrepreneurship.
Q: Do you have any sales expectations for the new phone? And from a technical perspective, what’s the difference between your approach and the ‘Doubao Phone’?
Yin Qi: First, the Doubao phone is a very important exploration for the whole industry. It brought the concept of the agent phone into everyone’s view. But our definition of the agent phone, including our definition of Step AOS, has some very unique ideas.
Previous Agentic OSes were more like iterative upgrades on top of Android. We think that path is a bit different from ours. Essentially, this OS should be a layer that sits on top of all future terminal-side OSes. It builds a runtime environment for the agent from scratch, based on the principle of human-machine symbiosis. It runs compatibility layers under Android/Linux/RTOS.
We believe that an operating system like this should be cross-device, not trapped on just one phone or one PC. That’s our definition of the OS.
With this definition of the OS, we’re probably quite different from existing terminal makers, including Doubao, in how they define the Agent system. I think there’s no standard answer right now. The ideal scenario is for everyone to try things out quickly and well, and then see which path is more correct.
Q: Internet application companies have their own ecosystem moats. The Doubao phone also hit some big bottlenecks with that. How does Jieyue plan to solve this problem?
Yin Qi: First, this is a very important question. In the age of the agent phone, besides changes in hardware and major shifts in the OS, what’s even more critical is that the underlying business logic and ecosystem are changing. This change is the result of co-creation by all the roles in the ecosystem. So, we can’t define what the stable state will look like right now. That’s my first point.
My second point: during this co-creation process, the logos you saw on stage were our first batch of major partners. We had very good communication with these big companies. They are very embracing of the agent era. Our collaboration is through APIs, not GUIs. This gives us a very good security mechanism between each other. And we believe that some industry standards will form very quickly. This is something we all need to push forward together.
At the same time, I have a more long-term thought. In the future, this kind of network might be more efficient than the current internet. On this network, advertising might not disappear, but it won’t be the main format. Maybe in the future, everyone on the chain does their own service well, and a service-based revenue sharing model based on roles will be more efficient.
Q: I still don’t quite understand how AOS actually solves these permission issues. Is it at the software layer, or is it solved from the software/hardware Amoo perspective?
Yin Qi: First is the security issue. We now essentially believe the most important part is at the OS level, solving it with software. I mentioned earlier that traditional open-source systems have their own security systems. Our Step AOS fundamentally re-engineers the Android, Linux, and RTOS layers underneath and builds an Agent layer on top. The ultimate security here is data-based security. As I said, you need more context to make a judgment.
Of course, we have many mechanisms, like sandboxing and rollback—those are ‘hard’ mechanisms. The most critical security mechanism is a dynamic, data-based one. For example, if we want to delete a photo, it needs to understand the context: is this photo sensitive? Should it be deleted? This is all based on the contextual data layer. There are many complex mechanisms here. We have at least one of the largest dedicated security teams in our entire OS, responsible for the entire OS security and privacy system. There’s a huge amount of software architecture design involved.
Q: For example, the data—does it belong to Amoo, or to the software partners you’re working with?
Yin Qi: We need to look at this in layers. First, all data needs to be under the premise of compliance and management. I believe the data ultimately belongs to the user.
At the same time, when our Jieyue Amoo interacts with Meituan’s API or other agents, there should be data boundaries. The data should respect the existing data boundaries of the original internet applications.
Q: Have you already negotiated those boundaries, or are they still not settled?
Ni Jiayue (President of Jieyue Terminal): Our first batch of ecosystem partnerships are settled. I think you might be asking more about the underlying part. As I said, we started from the bottom, with an open operating system. The first phone we made involved a fundamental restructuring of Android’s underlying architecture. For our next product, say a Linux-based one, we’ll restructure Linux from the ground up.
Q: Making hardware involves a lot of complex engineering—supply chain, quality control, after-sales service. These are completely different dimensions from building large models. Jieyue’s core team doesn’t have a background in that. How did you fill that gap in hardware expertise?
Yin Qi: First, in my more than ten years of entrepreneurship, my background has always been a mix of software and hardware. My PhD was in hardware sensors. I genuinely am interested in both. Second, Ni Jiayue is a veteran in hardware terminals. We have a complete team.
Within the current Jieyue system, our parent company is a very pure large-model-driven company, a software-centric company. Under Jieyue, the Jieyue Terminal system has a very strong software-hardware integration DNA. I think the DNA has to be right from the start. If a company doesn’t have that software-hardware integration DNA, it’s very hard to do hardware well.
Q: On the model side, while pushing for deployment, top-tier model companies are also stepping up exploration in foundational models, including new directions like RSI. Do you have any thoughts or directions on new architectures or technical paths that you can share now? What’s the approximate investment ratio for your foundational model?
Yin Qi: The investment in our foundational model currently holds an absolute dominant position within the entire Jieyue structure. We’ve always been a company that takes foundational models very seriously. Our main product direction is the agent, so we might invest more in the multimodal domain than other model companies.
You can look forward to our Step4 model. Step4 will be out in the second half of this year. After that, we might have a Step4 Pro. Both models will show good performance in the fundamental intelligence ceiling of large models, including long context and RSI.
The real differentiation for large models will be driven by application data. Without a Super APP pulling your model, everyone’s models will eventually reach a similar level over time. I think if you look two to three years out, you’ll see that Jieyue has the most solid technical capabilities and the most steadfast commitment to its technical path among China’s foundational model companies.
Q: Regarding on-device deployment, the industry is starting to see some homogeneity. Everyone is optimizing computing power, architecture, and user experience, using resources to push deployment. Can our approach through the operating system and native hardware break this current state?
Yin Qi: First, you have to look at it from different dimensions. From the model dimension, I don’t believe in a pure on-device model system. The cloud needs a model, and you can share the computing architecture and data. My view is, ‘If you can’t do a big model well, you can’t do a small model well either.’
Second, I also believe that ‘without strong large model capabilities, it’s hard to talk about deploying models on-device.’ We might not be the absolute best at on-device model deployment, but we are one of the most serious and dedicated companies building agent terminals.
When we look at deploying models on-device, we see the system is missing a lot of things. For example, the entire computing architecture of the chips. Our current on-device chips are completely unable to support the future demands of large numbers of users for agent capabilities. Can we, in the next two to three years, see a truly good on-device chip computing architecture that coordinates well with the cloud? That’s something we hope to drive forward.
I feel that the agent terminal is still in its very early stages.
Q: What stage is the phone at now? Do you have a prototype?
Yin Qi: Yes, we have one. Whether we can sell it depends on various factors, but we’ll push it forward as soon as possible. For us, the important point is that when people get this agent phone, the user shouldn’t have to ask us how it’s different from existing phones. For our first batch of early users, the evaluation we’re aiming for isn’t ‘this is a perfect product.’ The key is that people really feel it’s a new species. That’s the most important thing.
Q: In the past, people saw Jieyue more as a large model company. Now that Jieyue is moving into systems and terminals, how do you hope the outside world will understand Jieyue in the future?
Yin Qi: Jieyue is still first and foremost a large model company. Its most important goal is to push the entire AGI technology forward. It has a very good partner called Jieyue Terminal. Jieyue Terminal is a company that needs to complete the integration of software, hardware, and the business loop.
