With computing power, models, and hardware now aligned, AI is moving from a battle of hundreds of models to practical implementation and deep integration.

Article author and source: Lei Technology AGI

If last year’s World Artificial Intelligence Conference (WAIC) was a benchmarking frenzy dominated by large model vendors, all competing to boast the largest parameters and longest context windows, then the 2026 WAIC and the High-Level Meeting on Global Governance of Artificial Intelligence, opening on July 17, marks a new chapter.

This year, WAIC's core narrative has shifted, with the official theme being "Co-Creating the Future with Intelligent Partners." In our view, this theme reflects a more pragmatic industry consensus: large models are transitioning from concept to reality. The focus is no longer solely on compute power and parameters, but increasingly on the practical deployment of Agent (intelligent agent) productivity.

According to the latest information released by Shanghai’s Commission of Economy and Informatization, this year’s conference will focus on cutting-edge topics such as world models, open-source agents, AI coding, and token economies. But what excites us even more are the groundbreaking products set to make their global debut. From cloud-based computing behemoths, to the world’s first agent-powered smartphone, and humanoid robots capable of entering factories—WAIC 2026 will be a showcase of AI transforming the physical world.

At this year's WAIC, Lei Technology (ID: leitech)'s AI new media platform, LeiTech AGI (leikejiagi), will send a reporting team to Shanghai for exclusive coverage, bringing you comprehensive and professional insights. Now, let’s take a preview of this year’s WAIC and explore the key industry trends and major new product launches you won’t want to miss.

AI hardware and software make their debut, with an unprecedented number of first-time products unveiled.

This is expected to be one of the biggest highlights in the consumer electronics sector at this year's WAIC.

Many smartphones marketed with AI as a selling point often feature AI capabilities limited to a few standalone apps, such as AI photo editing or AI voice assistants. However, a true "agent phone" centers on a fundamental rearchitecture of the system-level agent engine.

On one hand, it must be capable of cross-app operations. AI can use its visual capabilities to "understand" the user’s current screen content. For example, if a user simply says, “Order me a coffee from the same place I had yesterday afternoon,” a system-level agent can autonomously open a food delivery app, locate the past order, place the order, and complete the payment—unlike the past state where different apps on mobile devices were isolated and operated independently.

On the other hand, the intelligent agent phone should have the capability for collaborative computing allocation between device and cloud. Leveraging powerful cloud-based models and lightweight on-device models, it can automatically allocate computing resources based on task complexity, ensuring both privacy and security while maintaining fast response times.

Although no manufacturer has yet claimed this AI agent phone, Lei Technology believes it is most likely from one of two companies:

The AI phone jointly developed by ByteDance and Nubia. Shortly after its launch, the DouBao Assistant phone quickly gained widespread popularity—but also faced immediate bans from major apps. However, this agent-driven approach has disrupted traditional smartphone product concepts and sparked significant upheaval.

(Source: Lei Technology)

2. Honor’s Agent Phone. Honor is a pioneer in AI phones and the only manufacturer to have launched a "Robot Phone." Notably, Lei Technology has received an exclusive invitation from Honor to the "New Paradigm for the Physical World" forum held during WAIC. According to the invitation, the Robot Phone and physical AI are the focal points of this forum. Whether Honor will unveil the Agent Phone concept will be revealed第一时间 by Lei Technology.

(Source: Lei Technology)

For the launch of this AI agent phone, we will pay closer attention to whether consensus can be reached among AI manufacturers, internet service platforms, and consumers regarding privacy protection and interest coordination under regulatory oversight, while also clarifying the boundaries of current AI agent phones.

Beyond hardware, the AI era brings a completely new software foundation. According to previews, leading Chinese large model company Jiepao Star will unveil its new Agent operating system at this conference.

Traditional operating systems, such as Windows and Android, are built around files and applications, with users seeking appropriate tools to accomplish tasks based on their needs. In contrast, Step’s Agent OS is fundamentally driven by intent and tasks.

Imagine a future where the screens of your devices are no longer cluttered with countless icons, but instead feature a single interactive gateway. Users simply express their needs in natural language—such as asking to create a PowerPoint presentation for work—and the Agent operating system automatically launches a browser to retrieve data, invokes large models for summarization and analysis, and employs layout software to generate the document.

In fact, many current PC-based tools like Lobster and Agent can already achieve this to some extent. Step星辰 previously released StepAI Desktop Companion, a system-level desktop Agent product that connects directly with the PC’s operating system, enabling management of local files, internet access, and the execution of complex tasks. By evolving from an Agent tool to an Agent operating system, Step星辰 has taken a significant leap forward—we’re curious how it will reimagine traditional operating systems and what use cases it can currently address.

(Source: Jiepoin星辰)

As a pioneer in domestic multimodal large models, MiniMax will showcase its M3 multimodal large model at WAIC. Launched on MiniMax’s official website on June 1 and officially open-sourced on June 15, M3’s key advantage is its native multimodality, enabling direct, deep understanding of text, images, and video—without relying on model stitching. Additionally, MiniMax M3 excels in Agent capabilities and coding, delivering maximum productivity. At WAIC, MiniMax is expected to demonstrate the model’s capabilities through more real-world use cases.

(Source: MiniMax)

Additionally, leading AI brands such as Zhipu and SenseTime will exhibit at this year’s WAIC, showcasing their latest large model products. Zhipu recently launched GLM-5.2, a model emphasizing long-context capabilities, which ranks among the top three on the Artificial Analysis leaderboard alongside Anthropic and OpenAI, making it the highest-ranked open-source model currently. SenseTime is likely to focus more on world models, with an emphasis on embodied intelligence.

In addition, Baimian Intelligence, an AI company focused on edge-side models, will also attend this year’s WAIC. In May, Baimian Intelligence unveiled its new MiniCPM5-1B model and the edge-side multimodal large model MiniCPM-V 4.6, which can run on just 6GB of memory. This year, Baimian Intelligence is likely to showcase the latest real-world deployments of these edge-side models on actual hardware.

(Source: FaceWall Intelligence)

Lenovo, having successfully ridden the AI wave, is no longer just the traditional PC manufacturer it was once perceived to be—it is now a key player in the AI industry. The AI infrastructure for this year’s World Cup in the United States, Canada, and Mexico was largely coordinated by Lenovo. Of course, at this year’s WAIC, Lenovo’s primary product offerings will likely still be PCs, but they will place greater emphasis on their AI capabilities.

Lenovo’s AI PC portfolio spans a wide range, featuring products powered by Intel and AMD chip platforms, as well as Qualcomm Snapdragon and future NVIDIA platforms. In addition to equipping PCs with local computing power, Lenovo has invested significant effort into developing its built-in personal agent system, “Lenovo Xiao Tian,” to deliver comprehensive agent capabilities to users. According to Lenovo’s vision, Xiao Tian possesses capabilities similar to a lobster, able to leverage the PC’s various functions to assist users in completing complex tasks; once workflow automation is achieved, it can be transformed into tangible productivity.

Six Link Intelligence is a company specializing in intelligent terminals and full-stack AI infrastructure services. Its product lineup includes laptops, tablets, mini PCs, all-in-one PCs, and NAS devices. While these products resemble traditional devices in form, their design philosophy and use cases differ significantly.

(Source: Liulian Intelligence)

The Agent PC products under SixLink incorporate AI chips from brands such as Intel and AMD at the hardware level, placing greater emphasis on memory, video memory performance, and AI computing power compared to traditional processors. At the software level, they provide Agent tools like SixClaw, enabling users to delegate tasks to AI for understanding, breaking down, and execution.

Sixth Link serves not only individual consumers but also B2B clients, offering complete AI solutions in addition to its hardware and software products. Sixth Link previously demonstrated a case in which supply chain, finance, and other operations were fully entrusted to AI running on Sixth Link hardware devices, creating an intelligent business system capable of autonomously performing tasks such as order optimization and cost analysis. At this year’s WAIC, Sixth Link will unveil its latest Agent PC product and showcase additional AI solutions tailored for government and enterprise customers.

As a typical application-layer company, WPS's core competitive advantage has never been its proprietary foundational model, but rather how to integrate AI technology into the real workflows of its over 600 million users.

Last year, WPS launched its native office AI agent, WPS Lingxi. This year, WPS is expected to shift its focus to demonstrating how it enhances complex daily office workflows, showcasing cross-app data flow—such as using a single command for AI to read a lengthy PDF financial report, automatically extract key data into an Excel table, and then convert it into a PPT document with one click.

(Source: WPS)

Another key office platform, DingTalk, will also participate in WAIC this year. Recently, Alibaba fully integrated its three major Agent platforms—QoderWork, Wukong, and MuleRun. After the merger, the Agent capabilities of these three platforms will be consolidated, and the new platform will continue to be led by DingTalk’s CEO.

DingTalk's current focus is on cross-app workflow automation. In deeply integrated automation agent solutions like Wukong, the key breakthrough lies in tool invocation and API orchestration. In real enterprise scenarios, agents can accomplish tedious tasks that previously required manual intervention across four or five different software applications. At WAIC, DingTalk aims to demonstrate a systematic engineering capability that reassures enterprises they do not need to build complex agent architectures from scratch—by simply integrating with DingTalk, they can automatically and securely execute intricate business processes.

Similar to how WPS and DingTalk have focused on optimizing office workflows, NetEase Youdao this year is also prioritizing the integration of AI agents with productivity tools and hardware terminals. Youdao is likely to unveil China’s first open-source AI agent, Youdao Lobster, demonstrating its ability to take control of local PC operations across applications and enable automated office workflows. It may also introduce its latest AI education hardware powered by the “Zi Yue 4.0” multimodal large model. This strategy of combining local AI agents with hardware enables Youdao to carve out a unique market niche and capture greater market share.

If the models, agents, and operating systems discussed earlier represent AI's soft power, then the computing infrastructure is the physical foundation supporting it all. At this year's WAIC, the highlight of Huawei's booth was undoubtedly the latest breakthroughs in computing infrastructure.

The highly anticipated Huawei Atlas 950 cluster will make its in-person debut. As a powerhouse designed for massive-scale AI training and inference, the Atlas 950 features significant upgrades in underlying architecture, high-speed interconnectivity, and cluster energy efficiency, making it a critical tool for domestic large model manufacturers aiming to achieve higher parameters and stronger multimodal capabilities this year.

(Source: Huawei)

Alongside the Atlas 950, we introduce a revolutionary near-memory computing 3D chip. In traditional architectures, the memory wall has long constrained the efficiency of on-device inference for large models due to bandwidth bottlenecks and high power consumption. This 3D-stacked chip overcomes these limitations by vertically integrating compute and memory units in three dimensions, achieving exponential gains in data transfer bandwidth and significant reductions in power usage. This represents a major leap in semiconductor technology and lays a solid foundation for the smooth execution of future trillion-parameter intelligent large models on-device.

Embodied intelligence is experiencing a boom, with robots flourishing in diverse forms.

Last year, Digu Robot launched its development platform, the S600, featuring up to 560 TOPS of computing power and an efficient, flexible big-brain/small-brain architecture. This year, WAIC presents an excellent opportunity for Digu to showcase its technology, with plans to prominently display multiple real-world embodiments powered by the S600—such as humanoid robots, industrial robotic arms, and collaborative robots—demonstrating its real-world capabilities in visual understanding, physical computation, and complex task planning.

(Source: Sweet Potato)

Additionally, as a foundational supplier, Digua Robotics may also showcase integrated hardware and software solutions to lower barriers for developers and enterprise customers, enabling robotic products to effectively address real-world challenges and enhance productivity.

At this year’s WAIC, AgiLab has clearly announced the launch of multiple new products, with its product line comprehensively covering full-sized humanoid robots to quadruped robots. As a key player in the embodied AI industry, AgiLab has demonstrated outstanding mass production and delivery capabilities: as of June this year, the total cumulative output of fully assembled AgiLab robots across all series has reached 15,000 units—just under three months after reaching 10,000 units in March.

Moreover, the robots under Zhìyuán have already undergone real-world scenario validation. In simpler terms, Zhìyuán robots are now capable of “going to work” in factories. Recently, a fleet of Zhìyuán Spirit G2 robots deployed at Longqi Technology’s Nanchang mass-production facility and conducted a six-day live stream of operations across the entire production line. At this year’s WAIC, Zhìyuán is likely to bring a 1:1 replica of these complex industrial precision assembly and logistics handling scenarios to its booth, directly addressing external skepticism about whether robots can truly perform real-world tasks.

In the past couple of days, Unitree Technology received approval for its IPO on the STAR Market, officially becoming the first humanoid robot company listed on China’s A-share market. Unitree has emerged as a highly prominent robotics brand in recent years, gaining widespread recognition through appearances at major events such as the Spring Festival Gala. According to its prospectus, the average selling price of Unitree’s humanoid robots remains around RMB 168,000, with a gross profit margin as high as 62.9%. In other words, Unitree has largely moved beyond the phase of selling at a loss to gain market attention and has successfully established a viable commercialization pathway.

Interestingly, Unitree Robotics often surprises audiences at product launches and other events. Last year at WAIC, Unitree set up a robot arena for its machines, drawing significant viewer attention. As expected, Unitree’s booth at this year’s WAIC will likely feature a similar interactive experience. Additionally, Unitree’s proprietary industrial-grade embodied large model, UnifoLM-X1-0, has already undergone pilot deployment and testing in its own factory, successfully completing precise assembly tasks for joint motors.

Therefore, at this year’s WAIC exhibition, Unitree is likely to publicly debut for the first time a robot equipped with this large model, showcasing its real-world performance on an industrial production line, directly addressing the expectations of the capital market.

Da Xiao has chosen a uniquely distinctive path in the embodied intelligence industry. As a player emerging from SenseTime, Da Xiao is likely to comprehensively showcase its ecosystem advantages at WAIC 2026. Recently, Da Xiao open-sourced its next-generation unified embodied foundation model, ACE-Brain-0.5, which delivers a complete set of capabilities—including spatial perception, decision-making, and embodied interaction—while maintaining a small parameter scale. This is expected to be one of the highlights of the exhibition.

(Source: Daxiao)

At the same time, Da Xiao’s unique ACE embodied R&D model will also be demonstrated live. Specifically, real people will wear haptic gloves and head-mounted cameras while performing actual tasks in real-life and work environments. The collected visual, force, and 3D trajectory data will be fed into a data acquisition system to train the robot’s brain.

Since Daxiao’s business model focuses on providing a general-purpose brain rather than specializing in a single hardware product, it is expected to bring several hardware ecosystem partners to the WAIC event to showcase real-world commercial applications, such as inspection robotic dogs and robot-operated stores.

Galaxy General represents a highly practical and commercially promising wheeled dual-arm approach in the embodied intelligence sector. Unlike companies focused on perfecting bipedal balance, Galaxy General believes that, at this stage, making robots walk like humans does not generate core economic value—what customers truly care about is the robot’s ability to effectively perform tasks with its hands.

The core strength of Galaxy General Robotics lies in its embodied vision large model, which can real-time analyze an object’s geometry, material, and center of mass—even if it has never encountered the object before—and quickly calculate the optimal grasping pose. Whether it’s a plush toy, a glass cup, or an irregularly shaped item, the robot can successfully grasp it using its control algorithms, achieving a success rate among the highest in the industry.

At this year's WAIC, Galaxy General is likely to highlight successfully implemented cases, including flexible material handling in factory logistics—transporting materials from storage areas to machine tools—and in unmanned retail, where robots can accurately identify, pick, and sort items for packaging.

This year at WAIC, there are many embodied AI companies, with hundreds of robots set to be showcased. Due to space limitations, we cannot review them all. However, these products share a common focus: a stronger emphasis on commercialization and practical deployment—essentially, how they deliver tangible productivity improvements for customers.

Summary: From the battle of hundreds of models to AI taking root in practical applications

If last year's WAIC was a competition centered on technical specifications, then WAIC 2026 is a serious test of commercialization. Reviewing this year's developments in hardware and software, large models, and embodied AI, we can clearly see the transformations taking place in the AI industry.

First, there is an explosion of agents at the terminal and system levels. AI is evolving from a simple web chat interface into digital employees capable of taking control of systems and autonomously executing workflows. Whether it’s ByteDance and Nubia potentially launching the world’s first agent-powered smartphone, Stepwise星辰 attempting to disrupt traditional PC logic with an agent-based operating system, or WPS and DingTalk taking over complex enterprise business processes, all point to a new trend: the future software ecosystem will still be driven by human intent, but executed by agents.

Secondly, computing power is beginning to shift toward local solutions and energy efficiency. For large models to truly enter the physical world, they must address bottlenecks in computing power growth—cloud computing resources are not infinite. This year, Lenovo and SixLink Intelligence will showcase AI PCs featuring NPU computing power exceeding 40 TOPS and large-capacity memory, which are becoming standard on-device specifications. Meanwhile, Huawei’s debut of the Atlas 950 beast and its near-memory computing 3D-stacked chip aim to reshape the very model of computing power generation at the physical layer.

Moreover, embodied intelligence has become more pragmatic. Two or three years ago, robots were mostly focused on showcasing flashy capabilities, but now, “working in factories” has emerged as a unified path for各家机器人 to demonstrate their capabilities. Unitree Technologies, which has just been approved for its科创板 IPO, signifies that humanoid robots have moved beyond the phase of selling at a loss for publicity and are now proving their commercial viability with shipment volumes in the tens of thousands. Meanwhile, wheeled dual-arm robots from brands like Galaxy General no longer insist on mimicking human form, instead adopting more practical technical approaches. The mass production of ten thousand units by AgiBot and the ACE development model by Daxiao Robotics are accelerating the growth of the embodied intelligence industry.