No one can precisely predict the next big thing, but no one wants to miss any possibility.Article author and source: Light Cone Intelligence
Attending WAIC this year was a spiritual delight but a physical ordeal.
Inside the exhibition hall, giant ice blocks were随处可见, audiences fanned themselves with stacks of promotional flyers, and staff sat in a row in the corner with cameras and lenses set up. Everyone’s face bore signs of exhaustion.
The total exhibition area of this year’s event has surpassed 100,000 square meters for the first time, with the forum and exhibition held in separate venues. This grand exhibition reflects the booming surge of both major corporations, small businesses, and entrepreneurs all accelerating into the AI industry.
The times are shifting; some participants are entering, while others are stepping away.
Large models, which were already sidelined last year, were even harder to find at this year’s exhibition. In their place were ubiquitous agent products, long lines of “AI agent phones,” and constant promotions of “industry brains” and “digital employees” from every booth.
After finishing the exhibition, Light Cone Intelligence identified several clear signals from this year's WAIC:
Models are no longer worthy of a booth; the spotlight has shifted from parameters to use cases. C-end companies are busy developing agents to secure first-mover advantage, while B-end companies are focused on building agent platforms or leveraging Forward Deployed Engineers (FDEs) to uncover enterprise transformation opportunities.
AI agents are flourishing, and AI hardware is no less impressive—nearly every booth at the venue is promoting AI agents. The booths for DouBao phones and JieYue Xingchen’s AI agent phones are always surrounded by long queues.
Yet beneath the surface of this boom, homogenization has become an inescapable issue. Countless AIGC creative platforms and countless agents with the “Claw” suffix fill the space—glance at them, and it’s hard to spot any differences. True differentiation ultimately hinges on each platform’s unique accumulated ecosystem, data, and access points.
“There’s a lot of uncertainty, but everyone is moving forward while observing,” a staff member from an exhibitor told Guangzhui Intelligence.
No one can precisely predict where the next big thing will emerge, but no one wants to miss any potential opportunity.
Agents become the absolute stars, but what do they look like?
The wind of AI applications coming to life has finally blown genuinely among the people. The most direct impression at the venue is that users’ acceptance of AI products has significantly increased, giving rise to so many independent AI applications.
Compared to last year, the number of users we’ve attracted this year has doubled.
Among numerous consumer-focused sectors, AI health is one of the most crowded this year. At the WAIC event, Ant’s Afu attracted on-site attendees to sign up by offering a scale for just RMB 0.01 to aid weight loss. Other consumer-oriented AI health assistants include JD Health’s AI doctor “Dawei,” iFlytek’s “iFlytek XiaoYi,” and Baichuan Intelligence’s “Bai XiaoYi.”
Through real-time observation and integration with the healthcare and supply chain industries, this year’s AI health products have evolved from initial diagnostic assistance to more advanced capabilities. Last year, AI could only offer general “suggestions”; this year, users can chat with digital avatars of real hospital doctors and even have medications delivered.
Focusing on specific products, agents are the absolute star of this year's C-end products.
Tencent has developed so many agent products that they can’t fit on a single wall. In addition to WorkBuddy and QClaw launched earlier this year, there’s Ardot for designers, Mavis—a more AI-powered digital assistant—and Toast, which lets users build standalone apps from scratch. These rapidly emerging agents seem more like internal tools first refined in-house before being rolled out to the public.
The team behind Alibaba's "Miaowu Team Edition" informed LightCones Intelligence that the creation of Miaowu stemmed primarily from internal needs. For instance, the functionality to create apps was highly requested internally; as more employees began using it voluntarily, the decision was made to launch a consumer-facing product. Later, demand from enterprise clients—for features such as registration forms and demonstration tools—led to the development of the enterprise version.
Some agents have been developed into standalone apps based on location services, while others have been directly integrated with the company’s flagship products to drive paid conversions through agent capabilities.
Baidu is a typical example. Besides its in-house work assistant DuMate and AI-powered creative product Miaoda, GenFlow on the booth appears more like an offshoot of Baidu Netdisk. For instance, GenFlow can activate conversational features within Baidu Netdisk’s main interface and leverage its agent capabilities to help organize key video information, draft documents, and more. When asked about GenFlow’s future plans, staff told LightCone Intelligence that it will likely be integrated into Baidu Netdisk’s subscription-based paid services.
At the bottom center of the Baidu Netdisk app, "Genflow" has appeared.
But amid this "prosperity" of agents, repetition is also very evident.
Light Cone Intelligence visited all the AI booths at the event. Although the products were numerous and overwhelming, they ultimately centered around a few key areas:
Various agents ending in "Claw," self-named personal agents that help you "build" applications, image and video agents that take over AIGC content creation, and even agent boxes resembling Mac mini devices.
Taking individual agents as an example, multiple products display nearly identical demonstrations in use cases, with data analysis and report generation serving as their “signature features.” When asked about their differentiated advantages, “support for multiple models” and “having a proprietary skill library” have become nearly universal responses.
After exploring, I was more impressed by Light Cone Intelligence’s Loomy and FaceBot Intelligence’s agent—both internally integrate a “knowledge base” feature that automatically categorizes user-uploaded files by type and applies AI-based basic annotations, making them more user-friendly for professionals.
Lommy's "Library" interface
Except for a few products with clear completion and well-defined goals, most are still in the exploratory phase, searching for breakout demand.
When Light Cone Intelligence speaks with representatives from multiple manufacturers, they are almost always asked in return: “What new features do you think we could add?”
“Everyone is doing it, and everything has become standardized. It’s too easy to adopt others’ features—just look at someone’s new feature and implement it. In practice, the truly useful features all converge.” A staff member told LightCones Intelligence.
This response may seem positive, but it’s truly born of resignation. The current boom in C-end agents isn’t really about competing for scenario ownership—it’s fundamentally an arms race of features. Everyone is copying, and the useful features are all fairly similar. In the end, differentiation can only come from ecosystems, data, and access points—not the product itself.
After OpenClaw, in the "Year of Agent Commercialization," every company hopes to find the next breakout use case. This is the collective picture of C-end agents—lively, but everyone is crammed together, with no clear differentiation.
But it seems that prosperity should always go hand in hand with a bubble.
Go B2B and reap the rewards—FDE is rising.
If the C-end is characterized by a flourishing diversity, then the B-end is walking a increasingly clear path—packaging its capabilities and offering them to more enterprises for purchase. The simplest example is selling APIs, a model already proven by Anthropic.
As large models continue to evolve according to the scaling law and agent products surge in popularity, B-end companies are primarily refining their existing workflows.
In terms of enterprise agent platforms and products, most last year were still focused on optimizing for specific scenarios. This year, the concept of “digital employees” has become more popular—beyond traditional roles like finance, operations, and customer service, new AI employee roles such as HR and administration have been added, bringing companies closer to having all their essential functions covered.
Taking NetEase Intelligence Enterprise as an example, its enterprise-grade AI Agent platform, ClawHive, can help a small manufacturing company build applications tailored to needs such as product quoting, raw material procurement, executive daily reports, and meeting decisions—significantly reducing manual effort in data collection, organization, and analysis, while enhancing efficiency in process management, critical decision-making, and project progression.
In vertical industries, launching intelligent agents is no longer surprising. For example, according to Guangzhui Intelligence, employees at iFlytek revealed that this year they launched the “Spark Legal Super Agent” and established a standalone company. It targets legal professionals such as lawyers, streamlining the entire litigation process—including case consultation and acceptance, pre-trial preparation, filing, and asset preservation. AI can now directly handle tasks like evidence organization and verification, risk identification, and generating various legal documents. This is highly valuable for lawyers, as in practice, clients often struggle to clearly articulate their own demands.
Compared to C-end, B-end is all about execution. Unlike last year’s focus on cost reduction and efficiency improvement, this year’s agents can take on more complex tasks based on their understanding of work projects.
A customer of Siemens Industrial Agent shared real-world results at WAIC. One application involves remote control of equipment such as pressure vessels—previously requiring on-site operations, now directly orchestrated by AI, significantly reducing maintenance costs. Another involves scheduling maintenance for offshore wind farms, where AI must simultaneously predict turbine conditions and weather patterns to perform maintenance within optimal time windows, minimizing power loss from wind generation.
B2B customers with higher demands for task success rates and reliability still need to invest more effort in engineering details.
Regarding current technical bottlenecks in agents, staff from FaceWall Intelligence noted that the real-world experience of agents is built upon countless engineering details, and many aspects still require refinement—for instance, agent memory. Next, FaceWall Intelligence aims to enable agents to “proactively identify tasks”: “You’re overseeing the project, but you’ve stepped away—leave it to the AI to fill in the gaps.”
Beyond standard products, Guangzhui Intelligence has observed another significant shift in the B2B market: the FDE model has gained popularity. At WAIC, an increasing number of companies are promoting their support for FDE, focusing first on deeply understanding enterprise AI needs before developing customized solutions.
WPS employees from Kingsoft Office told LightCones Intelligence that they began planning for FDE in the second half of 2025: “We have already successfully implemented several use cases, such as HR, finance, and legal. Now we are expanding to help enterprises by delivering various AI capabilities tailored to their business scenarios, such as agents.”
What is the necessity of FDE? A staff member from Wallface Intelligence told Lightcone Intelligence, “A company must have a small team, even if it’s just one person, who understands both the business and the technology.” He further revealed that Wallface Intelligence is also preparing to launch this initiative.
It’s somewhat ironic. AI was originally created to expect multiple solutions to a single problem and to possess generalization capabilities, yet in B2B deployments, it has reverted to a highly customized, close-collaboration FDE model—identical to the on-site implementation and consulting-driven delivery of the SaaS era. No matter how powerful an AI model vendor may be, closing the final mile of delivery still requires people who understand both the business and AI.
Perhaps the FDE model is a temporary solution due to the current limitations in AI productization. But overall, if it’s user-friendly and sells well, it’s a sound strategy—model providers can only continue down this path for now.
Large models recede into the background, world models gain popularity, and hardware becomes the new battleground.
Returning to the foundational layer of AI applications—the large models—the changes this year remain noteworthy. This year, Moonshot AI once again offered only a simple booth for visitors to check in. However, beyond WAIC, the attention sparked by Kimi K3 with 2.8T parameters has reached Silicon Valley.
At the WAIC event, the excitement around world models rivals the initial frenzy surrounding large models.
More critical changes are underway: AI is no longer just generating content—it is beginning to understand how the world works and predict the outcomes of actions. We are confident that 2026 will be the year of world models,” said Fang Han, Chairman and CEO of Kunlun Tech.
Based on this assessment, Kunlun Tech has extended its previous "AI all-in" enthusiasm to the world model赛道. At this year’s WAIC, they unveiled Matrix-Game 3.5, an interactive world model designed for gaming scenarios. Skywork’s Chief Scientist, Cheng Yu, revealed at the forum that the 3.5 version underwent a unified iteration across three key areas—data, model, and inference—building upon Matrix 3.0.
Since the technological approach for world models has not yet fully converged, this has led to divergence and differentiation within the space—from foundational world models for embodied intelligence to models applied in gaming and AI social industries.
Focusing on other model branches, another clear trend this year is real-time video models, which are entering the entertainment sector.
“When models enter the real world, the key is not just recognizing sounds and visuals, but understanding the context in which this information occurs, and using that to determine when and how to respond,” said a representative from MoSi Intelligence at WAIC.
This year, more players have entered the real-time video model space, such as Aishi Technology and Shengshu Technology, which previously invested in AI video and have since expanded into AI gaming and digital human interactions. At WAIC, we also experienced many AI interactive applications:
The AI interactive display brought by Google is called the "Emotion Cloud"—it assesses emotions through facial recognition. During the demonstration by Light Cone Intelligence, one user was identified as "anxious," causing the cloud above their head to turn dark. The user then held a soft, fuzzy emotion ball while listening to soothing words from the AI.
Certain visual auxiliary hardware is used to enhance interaction quality. iFlytek demonstrated a transparent glass screen on-site, which can assist in real-world interaction scenarios such as translation and answering questions. Staff explained to LightCone Intelligence that the camera is equipped to identify specific speakers in noisy environments and avoid misidentifying other people’s speech, thus preventing cross-talk.
However, the undisputed interactive star of WAIC 2026 is the agent phone. This year, two companies must be mentioned—one is the DouBao phone, and the other is Jieyue’s self-developed AI agent phone.
The second-generation phone from DouBao and Nubia appeared on site. After hands-on experience with Guangzhui Intelligence, the impression was that ByteDance applied the same patience it used to develop the DouBao app to building the DouBao phone. Compared to the first-generation model, which was more simplistic and direct, the second generation is more cautious and restrained in permissions, yet offers more refined and stable features.
Jieyue Xingchen went even further—they developed their own smartphone, drawing long lines of people on-site. After trying it out, the phone’s interface is quite bold, opening directly to an agent interface that demonstrates features like helping the phone remember fragmented memories and acting as a smart agent.
An increasing number of AI companies are not only collaborating extensively on hardware but also choosing to develop their own products.
Taking Mianbi Intelligence as an example, their previously announced AI hardware, the Pinea Pi, officially launched mid-year. This is an AI-native edge development board designed for developers, offering a demo validation environment by integrating modalities such as lighting, voice, and vision, along with deep optimization between hardware algorithms and underlying models.
When asked why they chose the developer community, the Pinecone team told LightCones Intelligence that this is a scenario that can be quickly validated: compared horizontally with products like smart glasses or fitness bands, the aforementioned workflow may appear feature-rich, but in the field of observation, there are still many aspects that require validation.
As a company specializing in edge models,
Mianbi Intelligence's approach to hardware resembles the common challenges faced by AI companies and hardware partners today.
“For endpoint model companies, if we only focus on software and collaborate with other hardware manufacturers, our bargaining power is quite low. By building our own supply chain and developing hardware ourselves, we can gain greater overall control,” he told Guangzhui Intelligence.
As physical AI is about to become a reality, how can AI software companies collaborate with hardware manufacturers to gain more influence? Large model companies are accelerating their alignment on this issue.
The model is evolving, requirements remain unclear, and boundaries are being tested. For all participants, there is only one choice this year—just get moving.