Is WAIC2026 a turning point for large model development?Article author and source: Lei Technology
On July 17, 2026, the World Artificial Intelligence Conference (WAIC) officially opened at the Shanghai World Expo Center. As an annual flagship event in the AI industry, WAIC serves as a kind of “progress check” for the sector: model developers, chip manufacturers, consumer brands, and robotics companies all gather in Shanghai to showcase their most representative products and technologies of the year.
Image source: Lei Technology
To stay up to date with the latest developments in the AI industry, Leikeji (ID: Leikeji) sent a team to Shanghai this year to bring you firsthand insights into the AI sector.
As in previous years, large models remain central to this year’s WAIC: Kimi released the K3 model with 3 trillion parameters ahead of the WAIC opening; MindSpore continues integrating MiniCPM into edge devices to bring intelligence to endpoint hardware; and companies such as Step星辰, DouBao, and Honor are embedding models directly into smartphone operating systems, transforming AI from a standalone feature into a system-level assistant capable of invoking apps and completing tasks.
But according to Lei Technology, compared to previous WAIC events, this year's WAIC has shifted its focus:
At the Zhìyuán booth, the GO-2 embodied foundation model powers robots to unload and pack products on the production line, while Magic Atom’s Magic-VLA K02 handles box sealing and clothing organization on-site. Even guided tours at WAIC are managed by over 60 Zhìyuán robots... An increasing number of large models are stepping out of the digital world and moving toward the physical realm, advancing toward physical AI.
Image source: Leju Robotics
In 2023, when large model technology first exploded, the AI industry competed over which models had larger parameters or supported more formats; even in 2025, the competition still centered on model context length and reasoning capabilities. Yet, “overnight,” large models that once operated solely within apps and APIs have stepped out of chat windows and into the physical world—and even the terminology has shifted from “large model foundations” to “foundation models.”
In the view of Lei Technology, the shift from the "digital world" to the "real world" is the most noteworthy development at this year's WAIC.
AI is no longer competing based on model parameter size.
Moving from the "digital world" to the "real world," what changes first is actually how we view large models.
For example, in the past, when models were released, people were primarily concerned with parameter scale, context length, and benchmark rankings; when CogView demonstrated MiniCPM at WAIC, they also highlighted its performance relative to other models of similar size.
However, for consumers, these specifications are difficult to directly translate into real-world usage experiences. After all, users won’t change their habits of booking tickets, checking the weather, or creating spreadsheets just because a model has a longer context window; even if a smartphone has numerous AI features, if users still need to constantly switch between maps, payments, messaging, documents, and browsers, it cannot be called a true AI device—only a “phone with AI apps.”
At this point, the concept of AI agents emerged. Users simply need to tell the AI what they want to do; the model can not only understand the request but also autonomously invoke the corresponding services to complete the task on the user’s behalf. The intelligent agent phone showcased at this year’s WAIC is precisely about this:
Jieyue Xingchen has integrated its personal agent, Amoo, into the底层 of the STEPX Neo mobile operating system, while DouBao and Nubia continue advancing GUI agents, enabling models to complete cross-app tasks by understanding screen content and simulating user interactions... Outside of mobile phones, this industry-wide shift toward agentization is even more pronounced.
Image source: Step星辰
Taking past AIoT hardware as an example, many so-called AI devices back then were merely Android phones with customized systems, such as the Rabbit R1. However, the AI devices we’ve tested this year by Lei Technology already begin to exhibit agent-like capabilities.
Embodied intelligence is AI as an 'Agent' in the real world.
Some may wonder why we’re talking about AI agents in the era of embodied intelligence—but the relationship behind this is actually very straightforward:
If we divide the evolution of large models over the past few years into stages, the first stage focused on the model’s ability to understand and generate content, the second stage emphasized Agent capabilities, and the third stage is centered on embodied intelligence and world models.
In the era of embodied intelligence, foundation models must not only understand what users say but also possess an understanding of the real world and the ability to "respond autonomously," such as VLA. Fortunately, at WAIC 2026, we have already seen the practical implementation of embodied intelligence and world models:
AI companies such as Zhìyuán, Magic Atom, and Lèjù all demonstrated embodied intelligence applications on factory assembly lines at WAIC; Qiyuan went even further by using robots to provide guided tours directly at the WAIC venue. Undoubtedly, these “assembly line” tasks lack the visual appeal of robots dancing or boxing, but from an industry perspective, these mundane, repetitive tasks better demonstrate the value of embodied intelligence and world models compared to performative demonstrations.
Image source: Leju Robotics
Moreover, these mundane "assembly lines" better reflect the current state of embodied AI: compared to home environments, factory tasks are relatively fixed, and the assembly line model means workflows can be infinitely broken down, with clear criteria for task completion.
More importantly, these scenarios can provide large models with real data in a relatively controlled environment: robots first operate in settings with clearly defined rules, companies identify issues and accumulate data from actual operations, then use this data to further train the models, ultimately extending their capabilities to more complex tasks.
Image source: Lei Technology
Precisely for this reason, an increasing number of mainstream AI large model companies at this year's WAIC are discussing embodied intelligence and world models, emphasizing that their products are set to be deployed at scale in the millions. In the view of Lei Technology, “stepping off the screen and into the real world” has become the industry consensus for AI in 2026.
To bridge into the real world, large models need new capabilities.
Of course, embodied intelligence entering the real world also introduces new demands on the capabilities of large models.
In the past, for example, in terms of thinking and response speed, waiting a few extra seconds in scenarios like chatting or image generation typically wouldn’t lead to serious consequences. However, in real-world scenarios represented by embodied intelligence, every second of delay from large models could result in severe consequences:
On a factory assembly line, even a one-second delay can be magnified, impacting overall production efficiency; in autonomous driving scenarios, response speed directly affects user safety.
At this year’s WAIC, we’re also seeing an increasing number of model companies reviving the concept of “edge-side intelligence,” leveraging the advantages of local deployment to reduce response times for embodied AI models.
Taking Minimax, the only company among the six new AI "little tigers" specializing in edge-side intelligence, as an example, Minimax demonstrated as early as 2024 that small models on the edge can also possess "large capabilities." Unlike edge models from other companies, Minimax’s edge models are not simplified versions distilled from full-scale cloud models; instead, they are independent, natively designed edge models powered by the Edge AGI architecture.
Image source: Lei Technology
This native on-device capability enables Mianbi’s on-device models to deliver fast responses and true reasoning without relying on a network. As a result, while other companies are still researching how to reduce model response times to make them viable in real-world applications, Mianbi Intelligence has already achieved mass production and delivery of on-device agents for the automotive industry.
However, just like that "mental math joke," a model that truly aims to operate in the real world must also learn to understand changes in the real world.
World models and VLAs have become popular in the field of embodied intelligence because they integrate perception, understanding, reasoning, and execution.
The user says, “Move this box onto the shelf.” The model must first understand the instruction, then use the camera to locate the box and the shelf, and finally control the robotic arm to grasp and place it. This entire process requires seamless collaboration among the language model, vision model, and robotic control system—any disconnect in one step will affect the model’s final performance.
This interconnected workflow makes the "full-stack model" a shortcut for large models to enter the real world: VLAs connect vision, language, and action; world models simulate the potential outcomes of those actions; and edge models enable rapid on-site responses. The Xuri S600 chip demonstrated by Digua Robotics at WAIC is the ultimate embodiment of this full-stack approach at the hardware level.
Is WAIC 2026 the turning point for large models entering industries?
After visiting this year's WAIC, Xiao Lei found that the industry's competitive focus has already shifted—from models and terminals to robots.
While parameters, context, and reasoning capabilities are important, they primarily determine a model’s upper performance limit; what defines a model’s lower bound—its baseline reliability—is on-device deployment capability, world modeling ability, and full-stack competence, all of which customers increasingly prioritize in the era of embodied intelligence.
From this perspective, it is only natural that the theme of WAIC 2026 has changed: large model companies are discussing edge deployment and world models, robotics companies are focusing on technological implementation, and chip manufacturers are adjusting their product directions around robotics—these once relatively independent technological paths have now converged under the era of embodied intelligence.
Image source: Lei Technology
It is certain that the AI industry has reached a higher stage.
Le Technology believes that large models at this stage will move beyond their "student phase" of serving only consumer applications and begin to be deployed in more serious industry use cases. Facing new demands for greater efficiency, stability, and security, whoever can first complete this identity transformation may gain an early advantage in defining the industry and secure a more proactive position in the next phase of AI competition.
From the performances of various companies at WAIC 2026, China's AI industry is clearly ready.