Alibaba

ByteDance

CATL

Huawei

Meituan

Tencent

Xiaomi

Most Followed

KrASIA Connection

Venture Roundup

Earnings Reports

China's Auto Sector

Inside China's Startups

100 AI Creators

View All

Collections

News

Features

Insights

Ecosystem

Press Releases

DeepSeek

Li Auto

Luckin Coffee

Mixue Bingcheng

Pop Mart

Xpeng

On the Radar

Events

Latest

About

Consumer

Robotics

Fintech

Collaborate

Baidu veteran Shen Pengfei joins Lee Kai-Fu at 01.AI as co-founder

Artificial Intelligence

Deep Tech

China

… priority. Given fierce domestic competition, the company is shifting its self-branded business focus overseas, and its collaboration with ByteDance represents a key step in that plan.
Seres also intends to allocate 20% of the proceeds from its upcoming Hong Kong listing to marketing, overseas sales, and charging network expansion to strengthen global brand recognition.
In the years ahead, Seres’s dual identity may become more pronounced: a loyal Huawei partner in China, and an ambitious, independent contender in international markets.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This <a href="https://36kr.com/p/3423694660275841">article</a> was written by Fan Shuqi for 36Kr.

Seres turns to ByteDance’s Doubao as it trims dependence on Huawei

East Asia

Transportation

Electric Vehicles

Mobility

Foundation and Large Models

… operations.
<h3>Hubble raises JPY 800 million to expand AI-powered contract platform</h3>
Hubble, a Tokyo-based provider of cloud-based contract management software, has raised an additional JPY 800 million (USD 5.3 million) in Series B funding, completing the round. The latest tranche follows a third-party allotment of shares to Vertex Ventures Southeast Asia &amp; India and DNX Ventures, bringing Hubble’s total Series B funding to JPY 2.3 billion (USD 15.2 million).
The company will use the funds to enhance its AI-powered contract assistant, strengthen customer success operations, and expand internationally.
According to CEO Shinpei Hayakawa, Hubble aims to build essential digital infrastructure for contract management, emphasizing rapid feature development and long-term scalability.
<h3>UMe Tea raises USD 10 million Series A to expand US footprint</h3>
<a href="https://www.umetea.com/">UMe Tea</a>, a California-based tea chain, has raised USD 10 million in a Series A round led by Conductive Ventures, with participation from iFly.vc and Enlight Growth Partners, according to <a href="https://36kr.com/p/3521277005896841">36Kr</a>. The funds will support team expansion and new store openings across the San Francisco Bay Area.
Founded in 2019 by former Silicon Valley engineer Jiachun Li, UMe Tea combines elements of Chinese tea culture with flavors tailored for US consumers. The company currently operates 27 stores and plans to reach 33 by the end of 2025, focusing on high-density locations in California.
UMe Tea told 36Kr that non-Asian customers now account for 60% of its customer base. Signature items such as mango teas and salted popcorn chicken have performed strongly, while limited-edition launches and collectible campaigns have helped drive engagement and sales growth.
whereAdBanner.Uptake
<h3>Recent deals completed in China:</h3>
<ul>
<li aria-level="1"><a href="https://neolix.net/">Neolix</a>, an autonomous driving technology company, has raised more than USD 600 million in a Series D round led by UAE-based Stone Venture, with participation from Gaocheng Capital, CITIC Capital, CDH, Sparkedge Capital, Beijing AI Fund, and an undisclosed major Chinese internet platform. The funds will be used to advance algorithm upgrades, product development, commercialization, and global expansion.</li>
<li aria-level="1"><a href="https://www.bio-compass.ai/Home?_l=en">BioCompass</a> (Baiyao Technology), a developer of virtual cell platforms, has raised an undisclosed eight-figure RMB sum in an angel funding round led by FreesFund, with participation from Shunxi Capital. Mingde Capital served as the financial advisor. Founded by a team specializing in bioinformatics and AI, BioCompass will use the new funds to advance its graph-based pretraining models for applications in gene regulation and cell state prediction under different environmental conditions. —<a href="https://36kr.com/p/3519810822249601">36Kr</a></li>
<li aria-level="1">Guoyan New Power Technology, a Shenzhen-registered lithium battery manufacturer, has secured nearly RMB 50 million (USD 7 million) in a Series B round from Haiheng Group. Established in 2019, the company focuses on the R&amp;D, production, and sales of steel shell lithium batteries and has developed proprietary composite cap and encapsulation technologies to support manufacturing. The proceeds will be used to expand production capacity and invest in new product development. —<a href="https://36kr.com/p/3519819794176902">36Kr</a></li>
</ul>
whereRelatedArticle
<h3>Pave Bank, Endowus, Hangry, and more made recent headlines:</h3>
<ul>
<li aria-level="1">Pave Bank, a commercial bank, raised more than USD 39 million in a funding round led by Accel, bringing its total funding to over USD 44 million. The round drew participation from other investors including Tether Investments, Quona Capital, Wintermute, Helios Digital Ventures, Financial Technology Partners, Yolo Investments, Kazea Fund, and GC&amp;H Investments.</li>
<li aria-level="1">Endowus, a digital wealth advisory platform, raised more than USD 70 million in its latest funding round led by Illuminate Financial, with participation from Prosus Ventures, Citi Ventures, and several Asian family offices. The round includes a USD 20 million cash injection and the conversion of existing notes, bringing the company’s total funding to over USD 130 million to date.</li>
<li aria-level="1">Hangry, a Jakarta-based operator of “virtual” restaurants serving delivery-only orders, raised USD 10.5 million in a Series A5 round led by Alpha JWC Ventures.</li>
</ul>
If there are any news or updates you'd like us to feature, get in touch with us at: <a href="mailto:hello@kr-asia.com">hello@kr-asia.com</a>.
<style>.addcaption{font-size:14px!important;line-height:20px!important;color:inherit;padding-top:8px;border-top:1px solid #e8e8e8}</style>

Deals in brief: Applied Materials backs robotics firm Augmentus, Sedifly raises pre-seed funding, Grab invests in May Mobility, and more

Venture Capital

Private Equity

Investment

Merger &amp; Acquisition

Singapore

Japan

Southeast Asia

Europe

The integration of <a href="/industry/ai">artificial intelligence</a> into assisted driving systems is accelerating, reflected in the rising parameter counts of on-vehicle AI models that now approach the scale of mainstream large models.
Case in point: according to 36Kr, <a href="/company/xpeng">Xpeng Motors’</a> upcoming on-vehicle AI model will feature at least seven billion parameters. <a href="/company/li-auto">Li Auto</a> is expected to reach a similar scale once it begins installing its self-developed chip next year.
Xpeng’s on-vehicle AI model is distilled from its in-house, cloud-based world foundation model, a design meant to address the compute, memory, and bandwidth limitations of vehicle chips that prevent deployment of full-scale models.
In the second half of 2024, Xpeng began transitioning toward large-scale cloud models. The company is developing an ultra large autonomous driving model with at least 72 billion parameters, slated for official release next month.
During a tech sharing session in April, Xpeng disclosed that its cloud model is built on a large language model backbone trained on extensive multimodal driving data. It reportedly interprets visual information, performs reasoning, and generates driving actions.
Once training is complete in the cloud, Xpeng plans to use knowledge distillation to compress the model, preserving key capabilities while creating a smaller version that can operate on vehicles. This process, also used by <a href="/company/deepseek">DeepSeek</a>, reduces model size without significant performance loss.
On the hardware front, Xpeng began developing its Turing chip in 2020. The chip entered mass production in June and debuted in the Xpeng G7. Designed specifically for AI and end-to-end large model workloads, it delivers about 700 TOPS (tera operations per second) of computing power, comparable to Nvidia’s latest Thor chip, and can process models with up to 30 billion parameters.
In early August, Xpeng held an internal meeting chaired by CEO He Xiaopeng, directing all AI resources toward the foundation model team to support the rollout of the seven-billion-parameter on-vehicle model.
Li Auto has likewise ramped up its AI efforts.
During its second-quarter earnings call, CEO Li Xiang said the company’s current on-vehicle large model has more than four billion parameters, marking a tenfold increase over its previous end-to-end version. Several industry insiders told 36Kr that once Li Auto’s self-developed chip rolls out next year, its vision-language-action (VLA) model will also reach around seven billion parameters.
Initially, Li Auto deployed a smaller and slower vision-language model (VLM) on its vehicles. In October 2023, the company launched an assisted driving system combining end-to-end and VLM architectures. The end-to-end model served as the fast system, while the VLM acted as a slower copilot, each running on separate Orin X chips.
In this setup, the end-to-end system functioned as the driver’s primary “brain,” while the VLM offered secondary input, preventing the larger model from running at full capacity.
Now, Li Auto is shifting focus to the VLA framework, originally introduced by DeepMind and now widely adopted in embodied intelligence. Generally speaking, VLA architectures integrate vision, language, and action reasoning to mimic human cognition, enabling perception, understanding, and task execution. This approach is gaining traction among automakers, including Li Auto and Xpeng.
To accelerate VLA model deployment next year, Li Auto has restructured internally. In May, Xia Zhongpu, who led the company’s end-to-end assisted driving program, <a href="https://kr-asia.com/li-autos-e2e-architect-departs-as-company-shifts-focus-to-vla">departed</a>. In September, Li Auto reorganized its autonomous driving team into 11 sub-departments under a flatter structure to speed up model development.
Beyond Xpeng and Li Auto, <a href="/company/huawei">Huawei’s</a> WEWA architecture uses a cloud-based world engine to train its vehicle-side world model, while <a href="/company/nio">Nio</a> is also deploying large world models on its vehicles.
<a href="/company/tesla">Tesla</a>, often regarded as the industry leader in assisted driving, has achieved regional robotaxi capability using end-to-end technology rather than large models. Suppliers such as Horizon Robotics and Momenta have also delivered strong performance through similar end-to-end approaches.
Meanwhile, some automakers that heavily promote their AI strategies have been caught up—or even surpassed—in real-world assisted driving performance. This suggests that model size alone does not determine capability. End-to-end systems are effective at learning from human driving behavior, while larger models may offer advantages in reasoning and decision-making. However, the core of assisted driving remains spatial perception, where the benefits of large model reasoning are still limited.
For automakers still refining end-to-end systems, building ever-larger models can divert computing resources toward linguistic reasoning rather than perception, potentially degrading the overall driving experience.
Many automakers are now pursuing ambitions beyond vehicles. Some see themselves as players in the broader field of embodied intelligence, with Li Auto among them.
During a company live stream last December, CEO Li described Li Auto as an AI company that still makes cars but views vehicles as “spatial robots,” applying its broader AI vision to mobility.
Xpeng has similar ambitions. The company plans to use its Turing chip not only in vehicles but also in robotics and flying cars, extending its autonomous driving capabilities to other intelligent systems.
Marketing is another factor. Following the global attention generated by ChatGPT, “large models” has become a powerful industry buzzword, just as “end-to-end” gained traction after Tesla’s V12 rollout in North America. For some automakers, highlighting large model development serves as both a technological statement and a way to capture public attention.
Still, regardless of motivation, improving real-world assisted driving performance should remain the industry’s core priority. Narratives that stray from this focus risk losing sight of what ultimately matters to drivers and passengers alike.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This <a href="https://36kr.com/p/3510288150944643">article</a> was written by Fan Shuqi for 36Kr.

China’s automakers race toward large AI models for assisted driving

Xpeng Motors

Autonomous Vehicles

… noting that rising commodity prices in China have usually "heralded improving household willingness to spend more in the coming months.
"We expect such a correlation will persist, as rising prices should boost producers' margins and alleviate consumers' deflationary concerns," Hong wrote. "So, the producers will hire more and provide more jobs, consumers will earn more and spend more. Eventually, if everything falls into place, a virtuous price and growth cycle should rejuvenate," he added, suggesting the project's impact could amount to three to four times its RMB 1.2 trillion price tag.
Economists who have tried to pin down the project's potential impact more finely, such as Lu Ting of Nomura, Citigroup's Xiangrong Yu, and Leah Fahy of Capital Economics, forecast that Yarlung Tsangpo will add about 0.1% to China's annual gross domestic product in the project's early stages.
For some, that just is not enough to make Yarlung Tsangpo really matter for the country's economy.
"It is clearly not a game changer for China's overall growth," said Homin Lee, a senior macro strategist at Swiss private bank Lombard Odier. Added Larry Hu, China economist at Macquarie Group, "We don't share the view that the construction of a mega dam on the Yarlung Zangbo River is another step toward reflation."
Some analysts, such as Jeff Zhang of Morningstar, even question whether the commodity requirements of the project will be meaningful given the huge scale of the Chinese market.
"At first sight, an estimated four to six million tons of steel and 30–50 million tons of cement are huge volumes, but they are still not sufficient to make a difference in the markets, in particular not over an estimated ten-year construction period," wrote Carsten Menke, head of next generation research at Swiss private bank Julius Baer, in a research note.
For such critics, Yarlung Tsangpo is actually a sign of Beijing's inability to break from its tried-and-true formula of turning to infrastructure investment to stimulate the economy during times of stress, despite years of urging by economists for the authorities to <a title="" href="https://asia.nikkei.com/economy/why-china-s-consumption-boost-has-fallen-flat" target="_blank" rel="noopener">look to consumers for growth</a>.
"The Yarlung Tsangpo megadam project thus calls China's principal ability to stimulate the economy via infrastructure investments into question, as neither from an economic point of view nor from a construction materials point of view is it big enough to make a major difference," Menke said. "And if such a project cannot make a difference anymore, which project can?"
This article first appeared on <a href="https://asia.nikkei.com/spotlight/policy-asia/will-world-s-largest-dam-project-be-big-enough-for-china-s-economy">Nikkei Asia</a>. It has been republished here as part of 36Kr’s ongoing partnership with <a href="https://www.kr-asia.com/nikkei-teams-up-with-36kr-in-asia-tech-news-coverage/">Nikkei</a>.

Will the world's largest dam project be big enough for China's economy?

Energy

Unitree Robotics has unveiled its latest humanoid robot, the H2, a full-sized bipedal machine priced at USD 29,900. The model combines industrial-grade engineering with a refined exterior design that the company describes on its website as “streamlined elegance.”
The launch marks a shift in Unitree’s design language. The H2’s head has been reshaped to appear more humanlike, with smoother contours and a bionic facial structure integrated with a dual-eye camera system for stereo vision. According to the company, this redesign aims to balance form and function, positioning humanoid robots as both approachable and efficient.
Standing 1.82 meters tall and weighing about 70 kilograms, the H2 features 31 degrees of freedom: six per leg, seven per arm, and three at the waist. Each limb is powered by internal rotor PMSMs (permanent magnet synchronous motors), generating up to 120 newton-meters of torque in the arm joints and 360 newton-meters in the legs. The frame is built from aircraft-grade aluminum, titanium alloy, and high-strength engineering plastics to optimize the strength-to-weight ratio.
<figure id="attachment_157976" aria-describedby="caption-attachment-157976" style="width: 480px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" class="size-full wp-image-157976" src="https://console.kr-asia.com/wp-content/uploads/2025/10/Unitree-Robotics-H2-Movement.webp" alt="A snippet of the H2 humanoid robot’s launch video, showcasing its durability and dexterity through a series of combat-style movements." width="480" height="270" /><figcaption id="caption-attachment-157976" class="wp-caption-text">In its launch video, Unitree Robotics demonstrates the H2 robot’s durability and dexterity through a series of combat-style movements. Graphic and header image source: Unitree Robotics.</figcaption></figure>
Power comes from a lithium battery with a capacity of 15 ampere-hours, offering up to three hours of continuous operation. The H2 supports Wi-Fi 6 and Bluetooth 5.2, includes voice interaction via an array microphone and speaker, and offers modular computing options ranging from Intel Core i5 and i7 processors to <a href="/company/nvidia">Nvidia’s</a> Jetson AGX Thor, which delivers up to 2070 TOPS (tera operations per second) of <a href="/industry/ai">artificial intelligence</a> compute.
A variant, the H2 EDU, targets research and academic users. It supports secondary development, customizable compute modules, and includes a 12-month warranty, compared with eight months for the standard version.
Beyond its technical profile, the H2 reflects Unitree’s effort to redefine its <a href="/industry/robotics">robotics</a> aesthetic. The company’s “streamlined elegance” initiative emphasizes smoother and more humanlike exteriors, mirroring a broader industry trend to humanize robots designed for social or domestic settings. Whether consumers will favor humanlike designs or prefer more mechanical forms, however, remains uncertain as humanoid robots gradually approach commercial viability.
<figure id="attachment_157975" aria-describedby="caption-attachment-157975" style="width: 480px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" class="size-full wp-image-157975" src="https://console.kr-asia.com/wp-content/uploads/2025/10/Unitree-Robotics-H2-Catwalk.webp" alt="A catwalk sequence with Unitree's H2 robot dressed in costumes walking alongside human models." width="480" height="270" /><figcaption id="caption-attachment-157975" class="wp-caption-text">To emphasize the H2’s more humanlike design, the launch video features a catwalk sequence with the robot dressed in costumes, walking alongside human models. Graphic source: Unitree Robotics.</figcaption></figure>
At under USD 30,000, the H2 is among the most affordable humanoid robots on the market, significantly undercutting models from both domestic and international peers. The pricing aligns with Unitree’s <a href="https://kr-asia.com/from-sci-fi-to-factory-floors-humanoid-robots-at-the-crossroads-of-technology-capital-and-ambition">broader strategy of accessibility</a>, which has helped it capture an estimated 60% share of the global quadruped robot market through models such as the Go2.
Competition, however, is intensifying. In China, Noetix Robotics recently introduced Bumi, a smaller humanoid robot with limited mobility but a lower price of RMB 9,998 (USD 1,400), about one-third the cost of the H2. The expanding range of models underscores how developers are racing to define the next mass market robotics platform—one that combines dexterity, safety, and a design suited to coexist in human environments.

Unitree’s new H2 brings “streamlined elegance” to humanoid robotics

Embodied Intelligence

… Conference, Alibaba Cloud argued that measuring only public cloud invocation volume is “like seeing just the tip of the iceberg.” Many clients deploy open-source Qwen models privately on their own servers, the company said, so their usage isn’t captured in external data.
Alibaba Cloud instead measures total revenue, including GPU rentals, platform, and model services. By this metric, it claims the top spot.
To further clarify model strength, Alibaba suggested differentiating between providers that sell self-developed models and those that resell third-party ones, as well as between models used for complex reasoning and simpler labeling tasks.
whereRelatedArticle
A Frost &amp; Sullivan report published in September echoed this logic. Based on daily invocation shares of self-developed models in the first half of 2025, Alibaba’s Qwen ranked first with 17.7%, followed by ByteDance’s Doubao at 14.1% and DeepSeek at 10.3%.
The price war will not last forever. Large model developers are moving from competition on infrastructure and token volume to competing on performance, service depth, and user experience.
In 2024, price cuts drove token growth. By 2025, that strategy has lost momentum. Tokens are already cheap; businesses now care more about efficiency and outcomes.
For smaller model developers unable to compete on scale, differentiation has become critical.
Recent examples illustrate the trend. <a href="https://kr-asia.com/is-kimi-k2-strong-enough-to-spark-chinas-next-deepseek-moment">Moonshot AI’s K2 model</a>, optimized for coding tasks, saw rapid adoption thanks to its strong price-performance ratio. Z.ai’s “GLM Coding Plan” has also expanded concurrent programming capacity.
Ultimately, business competition in the AI cloud market follows a familiar rule: when you can’t win by your rival’s metrics, redefine the game by your own. And in today’s large model race, the only metric that truly matters is how many customers are willing to pay.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This <a href="https://36kr.com/p/3501828940815236">article</a> was written by Deng Yongyi for 36Kr.

Who leads China’s AI cloud race? Alibaba, ByteDance, and Baidu all stake a claim

Baidu

Alibaba Cloud

Enterprise Services

Cloud Solutions

AI Infrastructure

… initial public offerings have also helped boost their markets.
Hong Kong's IPO market is the largest in the world this year, led by flotations from battery giant <a href="/company/catl">Contemporary Amperex Technology</a> (CATL) and miner <a href="https://asia.nikkei.com/business/markets/zijin-gold-soars-over-68-in-world-s-second-largest-ipo-this-year" target="_blank" rel="noopener">Zijin Gold</a>. Hong Kong's dramatic recovery from a prolonged slump has been driven by a mix of secondary listings by Chinese companies pulling out of US markets, geopolitical uncertainty and strong investor appetite for China's emerging technology companies.
In May, the Hong Kong Exchange launched the "Technology Enterprise Channel" (TECH), a scheme designed to facilitate new listings by specialist technology and <a href="/industry/biotech">biotech</a> companies. It also offers a confidential filing option, where early-stage companies can mitigate risks arising from premature and prolonged disclosure of information on business strategies or proprietary technologies.
In September, India's market regulator instituted a range of reforms across multiple asset markets as the country works to attract investment while also strengthening protections for retail investors.
The Securities and Exchange Board of India (SEBI), meanwhile, has simplified rules for sovereign-backed funds and overseas retail funds. Investors seen as low risk will now have to provide fewer documents and meet lower compliance requirements to participate in India's equity and debt markets.
Such funds make up almost two-thirds of foreign investors in these markets, and the changes show that "the regulator is removing key hurdles and showing a clear intent to bring in long-term capital," said Feroze Azeez, joint CEO of wealth manager Anand Rathi Wealth. But they "will not flip the switch overnight," Azeez said. This point was underscored as foreign investors have withdrawn about USD 12 billion from the Indian stock market so far this year.
Despite booming IPO markets in certain countries, some Asian bourses are at risk of losing out to the world's largest equity market. "Singapore and other Asian bourses, by implementing market reforms, hope to encourage startups, including unicorns, to pursue IPOs in their home countries rather than listing in the US," said Nomura's Kitano.
The US is still the dominant market for startups, especially for bigger companies. While there are more than 1,600 unicorn companies globally, the US is home to around half of these. China, with the next largest share, has just 17%, followed by India with 5%, according to data from Crunchbase.
"As an investor, choice is something you want from the equity market. The more you can choose from, typically, the happier you are as an active investor," according to JP Morgan's Treves.
However, bourses are likely to face a challenging balancing act between improving the quality of listings and increasing their quantity. Moves to encourage IPOs by loosening governance standards could spell trouble.
This article first appeared on <a href="https://asia.nikkei.com/business/markets/trading-asia/asian-stock-markets-take-on-wall-street-with-slate-of-reforms">Nikkei Asia</a>. It has been republished here as part of 36Kr’s ongoing partnership with <a href="https://www.kr-asia.com/nikkei-teams-up-with-36kr-in-asia-tech-news-coverage/">Nikkei</a>.

Asian stock markets take on Wall Street with slate of reforms

Vietnam

Indonesia

Philippines

South Korea

Hong Kong

India

<a href="https://www.buttonsair.com/">Buttons</a>, a luxury tech brand, has launched its latest product, the Solemate smart audiovisual robot, marking its first hardware device powered by Hali, an <a href="/industry/ai">artificial intelligence</a> agent developed by Terminus.
First introduced in November last year, Hali has evolved from a digital assistant into what the company calls a “life collaborator,” equipped with spatial awareness and physical interaction capabilities.
<figure id="attachment_157930" aria-describedby="caption-attachment-157930" style="width: 1280px" class="wp-caption alignnone"><img loading="lazy" decoding="async" class="size-full wp-image-157930" src="https://console.kr-asia.com/wp-content/uploads/2025/10/Buttons-Solemate.jpg" alt="Image of Buttons’ new Solemate robot." width="1280" height="720" srcset="https://console.kr-asia.com/wp-content/uploads/2025/10/Buttons-Solemate.jpg 1280w, https://console.kr-asia.com/wp-content/uploads/2025/10/Buttons-Solemate-720x405.jpg 720w, https://console.kr-asia.com/wp-content/uploads/2025/10/Buttons-Solemate-1024x576.jpg 1024w, https://console.kr-asia.com/wp-content/uploads/2025/10/Buttons-Solemate-768x432.jpg 768w, https://console.kr-asia.com/wp-content/uploads/2025/10/Buttons-Solemate-480x270.jpg 480w" sizes="(max-width: 1280px) 100vw, 1280px" /><figcaption id="caption-attachment-157930" class="wp-caption-text">Image of Buttons’ new Solemate robot. Image and header image courtesy of the company.</figcaption></figure>
Hali is built on a three-dimensional semantic memory model that links information such as object locations, events, and retrieval tasks to spatial coordinates and environmental context. This design reportedly enhances the intuitiveness and accuracy of information recall while enabling the AI to reason through an internal world model.
As the core of user interaction, Hali can perceive and process user requests in real time. Unlike legacy systems that rely on wake words or app-based triggers, Hali interprets a user’s position, intent, and surroundings as contextual cues for engagement. In practice, intelligent services no longer need to be summoned, activating proactively and seamlessly through spatial understanding.
Operationally, Hali first uses its edge model to analyze user intent and decompose tasks. At the planning level, it coordinates resources, devices, and actions across space and time. It recognizes home layouts, user movement patterns, and environmental changes, then preprocesses and encrypts data locally through its edge chips. The system collaborates with multiple domain-specific agents via cloud scheduling to complete complex tasks.
A GPU server cluster at Terminus’ Xuzhou computing center handles large-scale collaborative computation, powered by a hybrid intelligent computing engine that dynamically manages task distribution. Meanwhile, Terminus’s cloud-based large model performs path planning to help the robot avoid obstacles and navigate accurately. Vision-language models identify targets, while large language models generate execution strategies that are sent back to the device for decision-making and response.
Supported by Terminus’s Xuzhou computing resources, these processes are completed within seconds. To address the challenges of integrating different chip architectures, the company’s AIoT (<a href="/industry/ai">artificial intelligence</a>-of-things) computing platform has created a unified abstraction layer that connects multiple operator libraries across architectures. It also provides an end-to-end toolchain compatible with various inference and training frameworks, allowing models to run efficiently on heterogeneous computing infrastructure and significantly improving both inference and training performance.
Shao Ling, chief AI officer and global president of Terminus, noted that specialized AI agents remain limited to their respective domains. “They lack the ability to transfer learning across fields or handle new types of problems. Most operate purely in the digital world, without spatial perception or direct interaction with the physical environment,” he said.
According to Shao, the next key step is to achieve a closed loop of perception, reasoning, and action in the physical world. “A truly general intelligent agent must perceive the geometric structure and dynamic changes of the three-dimensional world, reason about spatial relationships, physical laws, and causal chains of actions, and ultimately act safely and effectively to influence the real world,” he said.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This <a href="https://36kr.com/p/3517360668204167">article</a> was written by Huang Nan for 36Kr.

Buttons launches Solemate, a luxury robot powered by Terminus’ Hali AI agent