A research team from Peking University has unveiled EvoPhys-World, described as the world's first human-centric, scene-level, fully controllable 5D world model. Built on Moore Threads' domestically-produced computing infrastructure, the model pushes AI-generated worlds from the "watchable and navigable" stage to one that is "manipulable, deeply interactive, and self-evolving."

The EvoPhys team argues that the essence of a world model is the simulation of a hyperdimensional space. While 3D world models focus on spatial structure and 4D models add the dimension of time, a 5D world model must also understand the states of objects across different parallel universes — how different choices lead to different futures, and how those future simulations inform current decisions.

EvoPhys-World is built on a unified base model with two core forms: Model as World Engine and Model as World Policy. In the World Engine form, the model can construct continuously evolving scene-level world memories based on real environments and generate future imaginations. It supports arbitrary trajectory roaming within new scenes and simulates interactions with any object in the scene, including manipulation and movement.

In the World Policy form, the model can predict future world states and recommend optimal action strategies. This enables it to understand not just how the world evolves, but to pre-enact how an intelligent agent should act next, forming a complete loop from world generation to world manipulation.

Technically, EvoPhys-World employs a Latent Memory Pool for long-term spatiotemporal memory, using a spatiotemporal importance mechanism to select and compress key implicit states. It also features a parallel generation architecture with Unified Token Chunk output and a hybrid attention mechanism that performs unified state-action token parallel inference.

The underlying computing infrastructure is powered by Moore Threads' full-stack domestic GPU ecosystem, marking a significant milestone for China's homegrown AI computing capabilities. By relying entirely on domestically produced hardware, the project demonstrates that advanced world model research can be conducted without dependence on foreign GPU supply chains.

The project is hosted at evophys.com and represents a significant step forward in the world model research paradigm — moving AI from "seeing the world" to "changing the world" through understanding physical causality, interaction dynamics, and multi-future prediction and selection.

DJI has for the first time publicly detailed its handheld imaging technology roadmap ahead of the upcoming launch of the Pocket 4P, the Pro version of its wildly popular pocket gimbal camera series.

The Pocket series has emerged as a major growth driver for DJI beyond its core drone business, with global sales surpassing 10 million units. According to IDC data, the global handheld intelligent camera market reached 16.65 million units shipped in 2025, growing 83% year-over-year, with DJI commanding a 62% market share.

At the heart of the Pocket 4P upgrade is DJI's second-generation color science technology. DJI product manager Fang Xiaoming explained that the new color science aims to reproduce the key characteristics of film imaging — higher highlight latitude, more natural texture, and richer color transitions — while maintaining the usability of 10-bit encoding. This positions the Pocket 4P as what DJI calls "a pocket cinema camera for everyone."

The Pocket 4P features a dual-camera telephoto configuration, a significant evolution from the single-camera design of previous models. DJI's spokesperson Zhang Xiaonan revealed that the company's imaging journey began with consumer drones, where it iterated technologies including omnidirectional obstacle avoidance, three-axis gimbals, stable image transmission, and 4K cameras. These capabilities were progressively miniaturized for handheld use.

The company's handheld imaging history dates back to 2015 with the original Osmo, which pioneered the integrated handheld gimbal camera form factor. The key breakthrough came with Pocket 3, whose rotating screen design and Live Photo functionality aligned perfectly with social media trends when WeChat introduced Live Photo support, propelling the product to mainstream popularity.

Looking ahead, DJI identified two development paths for the Pocket series: pushing imaging boundaries through lens and sensor upgrades within form-factor constraints, and optimizing the end-to-end workflow for easier content creation. The company acknowledged that AI will play a growing role in helping ordinary users produce high-quality video content with lower barriers.

The Pocket 4P faces increasing competition. Insta360 has entered the market with its Luna series developed in partnership with Leica, while OPPO and vivo have both initiated handheld gimbal camera projects expected to launch in Q4 2026. Despite this, DJI maintains that its pricing strategy will not change in response to competitive pressure.

Huawei Cloud made a sweeping product announcement at its INSPIRE conference today, unveiling over a dozen new AI offerings spanning infrastructure, model services, enterprise agent platforms, and industry solutions.

Huawei Cloud CEO Zhou Yuefeng introduced the world's first full-stack embodied AI development platform, CloudRobo, along with a new AI-native cloud entry point called Zhiyuan, the Lingqu intelligent computing cluster, and layered agentic memory storage. This marked Zhou's first major industry-facing keynote since assuming the role in late November 2025.

Zhou articulated a new paradigm for the Agentic AI era: an infrastructure model comprising efficient token factories, continuous learning capabilities, converged general-purpose and intelligent computing scheduling, and secure autonomous operations. Huawei Cloud is positioning this as a "silicon-based black soil" strategy to nurture the next generation of AI applications.

On the infrastructure front, Huawei Cloud launched the AICS Lingqu intelligent computing cluster, which supports 100,000-card cluster scale with total computing power reaching 200 EFLOPS. The system reduces token generation latency to under 10 milliseconds and delivers 5 million tokens per second throughput per thousand cards, with 99.95% online service availability.

The company also introduced AMS Agentic Memory Storage, a high-capacity, layered storage solution purpose-built for AI agents. Leveraging NPU-direct CMS hardware with PB-scale memory space and KV Cache layered pooling, the system reduces inference costs while supporting day-long tasks, effectively breaking the memory bottleneck for sustained agent operation.

In model services, Huawei Cloud upgraded its ModelArts platform with enhanced model matrices, semantic routing, RLaaS (Reinforcement Learning as a Service), and hardware-level trusted environments. The platform now supports models from partners including Kimi, Zhipu AI, DeepSeek, and MiniMax alongside Huawei's own Pangu series.

On the enterprise agent front, Huawei Cloud launched CodeArts for AI-powered software development, which has already attracted over 100,000 developers since its commercial release, and AgentArts for enterprise agent lifecycle management. The company also debuted smart office products under the OfficeAce brand.

Huawei Cloud emphasized its differentiated approach of supporting both public cloud and hybrid cloud deployments, noting 1,037 consecutive days without major security incidents across its customer base. In a notable distinction from the original Chinese article, this English version has been condensed to highlight the most globally relevant product launches and strategic direction.

UBTECH Robotics, the Hong Kong-listed humanoid robot company known as "the first humanoid robot stock," has launched the global first full-size biomimetic humanoid robot for pre-sale on the JD.com platform, with over 1,200 units ordered within three days of listing.

The robot features two variants — a male version at 183cm and 42kg, and a female version at 168cm and 35.2kg. Both models offer 88 degrees of freedom across the entire body, with a battery life of 2 to 4 hours. The product is positioned as an emotional companion and home interaction robot.

According to the company, a deposit of RMB 3,000 is required for pre-order, with the presale period ending July 15. Notably, the deposit is fully refundable before the deadline. UBTECH has stated that the product's full details, including its design and feature set, will be officially unveiled at a launch event later this month.

The product listing has drawn significant consumer attention, with over 150,000 people in Beijing alone browsing the product page within the first three days. The company has explicitly stated that the robot is intended for adult users only, reflecting the sophisticated nature of the biomimetic technology and the emotional companionship positioning.

Industry observers have noted that the pre-order pricing structure — with a RMB 3,000 deposit — suggests a full purchase price potentially in the range of several hundred thousand RMB, comparable to mid-range automobiles in China. However, UBTECH has not yet disclosed the final retail price.

The overwhelming response to the pre-sale underscores the growing consumer appetite for advanced humanoid robots in China's domestic market, particularly for applications beyond industrial use cases. While UBTECH has previously focused on educational and commercial robotics, this consumer-oriented product represents a significant strategic expansion into the personal robotics segment.

The 88 degrees of freedom across the full body is particularly noteworthy — it represents a significant technical achievement in biomimetic motion control, enabling the robot to replicate human-like movements with unprecedented fidelity for a consumer product at this scale.

The battle for AI platform dominance among China's three largest internet companies has entered a critical phase, as ByteDance, Alibaba, and Tencent each pursue distinct strategies to capture the emerging AI agent ecosystem.

This week, reports surfaced that Tencent is testing a native AI agent within WeChat, with plans to begin compliance review this month. The news alone added approximately RMB 360 billion to Tencent's market capitalization in a single day. WeChat, with its 1.4 billion monthly active users and nine-year-old mini-program ecosystem, represents the most formidable distribution advantage in China's AI platform race.

Alibaba's Qwen, meanwhile, announced it is opening its Brand Agent and Skill ecosystem to enterprise partners, with Luckin Coffee, KFC, and Mixue joining as early adopters. Qwen is pursuing an open platform strategy, recruiting third-party brands to fill its execution layer with agent capabilities. The approach mirrors the transition from the search engine era — when brands paid for visibility — to an agent-driven model where brands purchase operational rights within AI platforms.

ByteDance's Doubao, which leads the AI-native app market with 345 million monthly active users according to QuestMobile data, is taking the opposite approach. Rather than recruiting external brands, ByteDance is integrating its existing Douyin commerce infrastructure into Doubao's execution layer, creating a closed loop from content discovery to transaction completion entirely within its ecosystem.

The strategic divergence reflects each company's core strengths. Tencent leverages WeChat's unparalleled social graph and mini-program infrastructure as a foundational layer that all AI agents can be built upon. Alibaba relies on its deep e-commerce and local services ecosystem. ByteDance capitalizes on its content-driven commerce engine and massive user scale.

Meituan CEO Wang Xing articulated the emerging paradigm during the company's earnings call, noting that going forward, serving AI agents (ToA) is becoming as important as serving consumers (ToC) and merchants (ToB). Meituan chose to make its delivery and local services capabilities available as infrastructure for any AI platform to call upon, rather than building its own AI entry point.

As WeChat prepares to insert its AI layer into China's most ubiquitous communication platform, the window for other players to establish user habits before the landscape resettles is rapidly narrowing.

BYD Secretly Develops Humanoid Robot Codename 'Yao-Shun-Yu' as Auto Giants Race Into Embodied AI

BYD, China's largest electric vehicle manufacturer, has confirmed it is secretly developing humanoid robots under a project codenamed "Yao-Shun-Yu." The revelation came from BYD Executive Vice President Li Ke in a recent interview, shedding light on the automaker's ambitions beyond electric vehicles and into the rapidly emerging field of embodied AI.

The project was initiated in 2022 and operates under BYD's 15th Business Unit, which focuses on electronic integration and intelligence. Li Ke highlighted BYD's unique advantages in developing humanoid robots, noting the company's deep expertise in motors, batteries, electronics, precision manufacturing, and chips. The automaker also boasts a 4,000-plus engineer autonomous driving team and plans to invest 100 billion RMB in AI and automotive intelligence initiatives.

"The fundamental challenge in this space is that China's robots lack a brain, while US robots have strong brains but weak limbs," Li Ke said. "BYD aims to produce robots that excel in both dimensions." This dual focus on hardware and intelligence differentiates BYD's approach from many competitors who prioritize one over the other.

The company is taking a pragmatic approach to production. The final humanoid robot product may not be entirely self-manufactured; BYD could adopt an open platform strategy, allowing third-party components and software integration. This mirrors BYD's "technology fish pond" corporate strategy, where the company prepares a wide range of technologies in advance and deploys them when market conditions are right.

BYD would likely serve as its own largest customer for the robots. Potential use cases include deploying the humanoids as store greeters at BYD's expanding retail network and in factory floor roles across its manufacturing facilities. This captive demand provides a concrete revenue pathway that many robotics startups lack.

BYD is far from alone in this race. Tesla has begun mass production of its Optimus Gen-3 humanoid robot, with 50 units already deployed at its Shanghai factory. XPeng plans to mass produce its IRON humanoid by the end of 2026, and Li Auto has its own internal project codenamed Nexus. The convergence of automakers into humanoid robotics reflects a broader recognition that the manufacturing and AI capabilities developed for autonomous vehicles transfer naturally to general-purpose robots.

As China's largest EV maker doubles down on embodied AI, the distinction between automaker and robotics company continues to blur, with BYD positioning itself at the intersection of both worlds.

Chinese Company Kuawei Intelligence Tops WorldArena Global Benchmark in Embodied World Models

Kuawei Intelligence, a Chinese embodied AI company, has achieved the top ranking in the WorldArena Track 2 (Data Engine) global benchmark for May 2026, surpassing international competitors including WoW and BLM. The achievement marks a significant milestone for China's embodied AI sector and underscores the country's growing competitiveness in world model research.

The company's winning model, DSCFuncWorld, is built on the DexWorldModel architecture and specializes in generating high-quality robot training data. WorldArena's Track 2 is notable for its practical evaluation methodology: rather than measuring visual quality metrics, it assesses real robot task success rates, making it a more meaningful benchmark for real-world deployment readiness.

This latest triumph follows a series of significant achievements for Kuawei Intelligence. The company previously won the ICRA championship in the United States, where its dual-arm robot completed fine manipulation operations without any human intervention. It also topped the RoboTwin benchmark with a 94 percent average success rate across diverse manipulation tasks.

Kuawei's technical approach differs from conventional generative world models. The company uses DINO semantic feature space for state prediction rather than pixel-level generation, resulting in more robust and computationally efficient training data synthesis. The company has also open-sourced its EmbodiChain data infrastructure, contributing to the broader embodied AI ecosystem's development.

The real-world impact of Kuawei's technology is already evident across multiple industries. The company has been deployed in over 50 industries across more than 1,000 projects. On Hisense's production line, the system achieved a 99.99 percent success rate. At Midea's factory, its robotic sorting system operates at three times human efficiency. These deployments demonstrate that Kuawei's world models translate effectively from benchmarks to factory floors.

The company generated over 100 million RMB in revenue in 2024, signaling commercial viability in a sector often criticized for being long on promise and short on revenue. Its W1 Pro humanoid robot has shipped over 100 units to customers. Kuawei's client list reads like a who's who of Chinese manufacturing: BYD, GAC, Zoomlion, SANY, and Panasonic.

As embodied world models become the foundation for robotic training at scale, Kuawei Intelligence's benchmark-topping performance positions it as a key player in the global race to build robots that can understand, predict, and interact with the physical world as effectively as AI systems already navigate the digital one.

Zhejiang University Team Creates Visual Reasoning System That Lets Robots 'Think With Their Eyes' — 22x Faster Than Text

Researchers at Zhejiang University, in collaboration with Cornell University, the National University of Singapore, and Xidian University, have developed a breakthrough visual reasoning system that enables robots to "think with their eyes" rather than processing language-based internal monologues. The system, called VisualThink-VLA, achieves a 22.8x speed improvement over text-based reasoning approaches while also delivering higher accuracy.

The fundamental insight behind VisualThink-VLA is that traditional Vision-Language-Action (VLA) models rely on text-based chain-of-thought reasoning, where the robot essentially writes an internal essay describing each step before acting. This process takes an average of 8.377 seconds per step — painfully slow for real-time manipulation tasks. VisualThink-VLA replaces text tokens with visual reasoning tokens, reducing processing time to just 0.367 seconds per step.

The system employs a four-channel visual evidence architecture comprising Bounding Box, Edge, Motion, and Relation channels. Rather than using all four channels indiscriminately, VisualThink-VLA features an adaptive routing mechanism that selects only 2.22 channels per step on average, optimizing the balance between computational efficiency and reasoning quality.

Testing across eight benchmarks yielded a 92.63 percent average success rate, outperforming the text-based ECoT approach which achieved 85.09 percent. The speed advantage is even more pronounced: 22.8x faster while being more accurate — a rare combination in AI systems where speed and quality are typically traded off against each other.

The researchers validated the system on a PIPER NERO 7-degree-of-freedom robotic arm, demonstrating success in multi-object pick-and-place operations, relation-sensitive placement where object spatial relationships matter, contact-sensitive reorientation, and two-stage compound tasks that require sequential reasoning. The training data, dubbed "VisualEvidence-Set," contains 754,700 instructions covering diverse manipulation scenarios.

A key design advantage is that VisualThink-VLA operates as a plug-and-play module for existing VLA systems. This means robots currently using text-based reasoning can be upgraded without entirely replacing their underlying architecture. The paper is available on arXiv under identifier 2605.30011.

The work represents a paradigm shift from "write an essay then act" to "see-think-act," moving robot reasoning closer to how humans naturally operate — processing visual information directly rather than translating it through language. As robots are deployed in increasingly dynamic environments, the ability to reason visually at near-instant speeds could be a critical enabler for widespread adoption.

OEM Self-Developed Chips Reshape Auto Supply Chain: BYD, NIO, XPeng Challenge Third-Party Suppliers

At least five major automakers are now developing their own autonomous driving chips, signaling a tectonic shift in the automotive supply chain that threatens to upend the dominance of third-party chip suppliers. BYD, NIO, XPeng, Li Auto, and Tesla have all introduced or announced in-house chips, each claiming performance metrics that rival or surpass commercially available solutions.

BYD's Xuanji A3 chip, built on a 4nm process with 700 TOPS of computing power, entered mass production in May 2026. XPeng's Turing chip, delivering 750 TOPS, began production in the third quarter of 2025 with an annual output of approximately 1 million units. NIO's Shenji NX9031, fabricated on a 5nm process and exceeding 1,000 TOPS, has already shipped 150,000 units. Li Auto's Mach M100, also on 5nm, claims 1,280 TOPS of "effective compute." Tesla's HW4 rounds out the competitive landscape.

The impact on third-party suppliers has been immediate. Horizon Robotics, a leading Chinese autonomous driving chip company, saw its stock drop 7 percent following BYD's chip announcement. The decline is significant because BYD is Horizon's largest customer, accounting for roughly 50 percent of its revenue. When a dominant customer becomes a competitor, the business model faces existential risk.

However, Horizon Robotics maintains substantial defensive moats. The company's toolchain maturity is widely recognized as best-in-class, with 27 original equipment manufacturers (OEMs) and 42 brands integrated into its ecosystem. Horizon is also expanding its product line with the Star series, a cockpit-driving fusion chip that combines infotainment and autonomous driving capabilities in a single SoC — an area where most OEM chips have yet to compete.

The supply chain disruption extends beyond chipmakers to Tier 1 suppliers, who are being squeezed as OEMs take back system-level control over electronic architectures. In the short term, spanning 2026 to 2027, OEM-developed chips will likely remain confined to high-end models, leaving basic third-party chip business largely unaffected. The medium-term outlook from 2027 to 2028 becomes critical: if OEM chips successfully penetrate sub-200,000 RMB models at million-unit scale, the competitive dynamics shift decisively.

Looking further ahead, if in-house chip capability becomes a core competitive barrier for OEMs, third-party chip companies face systemic customer base shrinkage. For an industry built on the horizontal division of labor between automakers and suppliers, the vertical integration trend represents the most profound restructuring in a generation. The question is no longer whether OEMs can build chips, but what role remains for those who only sell them.

Once in 20 Years: China's Semiconductor Industry Surges as AI Reshapes the Memory Market

The global semiconductor industry is on the cusp of a historic milestone, with the market approaching the $1 trillion threshold. According to the World Semiconductor Trade Statistics (WSTS), the global semiconductor market reached $791.7 billion in 2025 and is projected to grow to $975 billion in 2026. This surge is unprecedented in the industry's history, marking what many analysts are calling a once-in-20-years transformation.

China's semiconductor sector is riding this wave with remarkable momentum. The country's integrated circuit (IC) exports surged 83.7 percent year-over-year to $103.5 billion in the first four months of 2026, reflecting the rapid expansion of domestic chip manufacturing capabilities. This growth comes as the memory market, a cornerstone of the semiconductor industry, undergoes a radical restructuring driven by artificial intelligence demand.

The memory market is projected to reach $551.6 billion in 2026, representing a staggering 134 percent increase. DRAM prices have climbed 10-fold since 2025, fueled by AI-driven demand for High Bandwidth Memory (HBM) used in data center AI accelerators. Server storage as a share of total memory demand has risen from 20 percent to 40 percent, signaling a fundamental shift away from the consumer electronics-driven model that defined the industry for decades.

The market transformation is reflected in soaring stock prices: SK Hynix has seen a 10x increase, while SanDisk has surged 30x as investors bet on memory-driven growth. In China, CXMT (ChangXin Memory Technologies) has disclosed its IPO prospectus, showcasing strong profitability and positioning itself as a serious player in the global DRAM market. YMTC (Yangtze Memory Technologies) is widely viewed as the next major contender likely to follow with a public listing.

Academicians are taking note of the shift. Xu Hongxing, a member of the Chinese Academy of Sciences, stated that China's chip industry has emerged from its toughest period and is now entering a phase of accelerated growth. The Chinese memory market alone is expected to reach approximately 1 trillion RMB, driven by downstream demand from AI servers, smart vehicles, and industrial robots.

The industry is in the midst of a structural transition, moving away from heavy reliance on consumer electronics toward AI-driven applications. Smart cars and robots are becoming the primary demand drivers, reshaping the competitive landscape. With domestic champions like CXMT and YMTC gaining ground, China's semiconductor resurgence appears to be more than a cyclical rebound — it signals a long-term strategic shift in the global balance of chip manufacturing power.

Huawei Unveils Mate 90 Series with Revolutionary Kirin 9050 'Logic Folding' Chip

Huawei has announced its highly anticipated Mate 90 series, slated for a September 2026 launch, positioning the flagship lineup to go head-to-head with Apple's iPhone 18 Pro. The centerpiece of the new devices is the Kirin 9050 processor, codenamed "Kirin 2026," which debuts an industry-first technology called "Logic Folding" — a 3D stacked transistor design that breaks away from traditional single-layer circuit architectures.

The Logic Folding innovation represents a fundamental rethinking of chip design. Unlike conventional planar transistor layouts, which arrange circuits in a single two-dimensional layer, the Kirin 9050 employs a three-dimensional stacking approach that layers transistors vertically. This architecture dramatically increases transistor density while reducing power consumption and heat generation, enabling performance gains that traditional node shrinks alone could no longer deliver.

Industry observers note that the Logic Folding technology could be Huawei's answer to the advanced process node restrictions it has faced due to US export controls. By innovating at the architecture level rather than relying solely on lithography advancements, Huawei has found a path to continue pushing chip performance forward despite limitations on accessing leading-edge fabrication equipment.

Beyond the chipset, the Mate 90 series promises substantial upgrades across multiple dimensions. The camera system has been overhauled with improved sensors and computational photography capabilities, building on Huawei's long-standing reputation for mobile imaging excellence. Communications hardware has also received significant enhancements, including improved satellite connectivity support and better overall signal performance. Huawei has also emphasized system smoothness, suggesting deep software-hardware co-optimization that leverages the Kirin 9050's unique architecture.

The timing of the Mate 90 launch puts it in direct competition with Apple's iPhone 18 Pro, expected around the same timeframe. Huawei's return to the high-end smartphone segment has been carefully orchestrated following the Mate 60 series' surprise success in 2023, which demonstrated that Chinese consumers remained enthusiastic about Huawei's flagship phones despite the absence of 5G capabilities at the time.

With the Kirin 9050 and its Logic Folding technology, Huawei is signaling that it intends to compete not just on brand loyalty but on genuine technological innovation. If the 3D stacked transistor design delivers on its promise, the Mate 90 series could mark a turning point in the smartphone silicon wars, proving that architectural breakthroughs can compensate for fabrication disadvantages.

Huawei has not yet disclosed full specifications or pricing, but the company has confirmed that the Mate 90 series will launch in September 2026 across multiple markets.

Tags: Huawei, Mate 90, Kirin Chip

China Achieves Mass Production Breakthrough with 360TB Glass Hard Drives

Researchers at Huazhong University of Science and Technology (HUST) have achieved small-scale mass production of glass-based hard drives, a breakthrough that could transform enterprise cold data storage. Each glass disc can store a staggering 360 terabytes of data across 400 stacked layers, using laser "carving" technology that writes data into the internal structure of the glass medium.

The technology, developed in collaboration with Wuhan-based startup YiYao Technology, uses femtosecond laser pulses to create microscopic modifications within glass discs, effectively encoding data in three dimensions. The 400-layer stacking capability represents a quantum leap in storage density compared to traditional magnetic hard drives or even solid-state drives.

Performance specifications reveal both the technology's strengths and current limitations. Write speeds range between 8 and 10 MB/s, while read speeds reach 50 to 200 MB/s. The drives are write-once media — data cannot be erased or rewritten once stored — making them unsuitable for active storage workloads but ideal for archival and cold storage applications.

YiYao Technology was founded in Wuhan and has attracted top talent from the global optical storage community. Notably, a former chief researcher from Microsoft's Project Silica has joined the company as co-founder, bringing invaluable expertise in glass-based data storage — a field Microsoft has researched for years but has yet to commercialize at scale.

The target market for these glass drives is enterprise cold data storage, a segment currently dominated by magnetic tape. Tape storage, while inexpensive, suffers from slow access times, mechanical degradation, and limited lifespan. Glass storage offers several compelling advantages: exceptional durability (glass discs are resistant to water, electromagnetic fields, and extreme temperatures), extremely long data retention measured in centuries rather than decades, and higher storage density per physical volume.

"If you think about data centers that need to store petabytes of archival data for regulatory compliance or historical preservation, glass storage is a game-changer," said an industry expert familiar with the technology. "The write-once nature is actually a feature for cold storage — it guarantees data integrity over time without risk of accidental deletion or corruption."

The HUST and YiYao team are now working to scale production volumes and improve write speeds. While the current 8-10 MB/s write rate is acceptable for archival workflows, faster writing would open additional use cases. The long-term vision includes competing not just with tape but with traditional hard drives for certain nearline storage applications.

China's glass storage breakthrough represents a rare convergence of academic research, industry talent, and manufacturing capability — and it positions YiYao Technology at the forefront of what could be the next generation of data storage infrastructure.

Tags: Glass Storage, HUST, Data Storage

ByteDance's Doubao Introduces Paid Pro Tier as User Base Dips by 6 Million

ByteDance's AI assistant Doubao has officially launched a paid subscription tier, the Doubao Pro version, with monthly pricing ranging from 68 to 500 RMB ($9 to $70). The move marks a significant shift for China's most popular AI chatbot by user count, but comes at a delicate moment as third-party data reveals the platform lost 6.07 million monthly active users in May — a 1.81% decline that brings its MAU to approximately 330 million.

Doubao has been a dominant force in China's consumer AI space, processing an eye-popping 120 trillion tokens daily. Its scale has made it the benchmark against which other Chinese AI assistants measure themselves. However, the introduction of a paid tier represents ByteDance's acknowledgment that even massive user numbers must eventually translate into revenue.

The pricing structure is notably aggressive. The entry-level 68 RMB per month tier places Doubao Pro above several competitors' basic subscription fees, while the 500 RMB premium tier targets power users who require higher usage limits and priority access. ByteDance is betting that Doubao's deeply integrated ecosystem — spanning its TikTok-like platform, productivity tools, and enterprise services — will provide enough added value to justify the cost.

But the user decline in May raises questions about timing. According to data from third-party analytics firms, Doubao's 1.81% MAU contraction occurred while several competitors posted growth. Qwen, DeepSeek, Yuanbao, and Kimi all saw their user bases expand during the same period, suggesting that Doubao may be losing ground even as it attempts to monetize.

The competitive landscape is intensifying rapidly. Alibaba's Qwen has been aggressively pushing both its open-source models and consumer applications. DeepSeek's astonishing cost efficiency has made it a favorite among developers. Tencent's Yuanbao and Moonshot AI's Kimi have also carved out loyal user segments. Doubao's challenge is to convince free users to pay without driving them to these alternatives.

ByteDance's calculus may be informed by its enormous capital expenditure plans. Reports indicate the company could spend up to $70 billion on AI infrastructure, making the need for revenue generation urgent. Doubao, as ByteDance's flagship AI consumer product, must contribute to that ROI.

Whether Doubao's paid tier succeeds will depend on how well ByteDance balances monetization with retention. The 6 million users lost in May represent a warning shot — and the coming months will reveal whether the paywall accelerates that trend or stabilizes the platform's finances.

Tags: ByteDance, Doubao, AI Subscription

Awinic Positions Itself as the 'Audio King' of AI Glasses with New Chip Solution

Shanghai-listed Awinic Technology (688798.SH) has unveiled its latest audio solution purpose-built for AI glasses, the AW88188, at the Songshan Lake IC Forum. The move positions the 18-year-old semiconductor company — which has shipped over 36 billion chips to date — as a dominant player in the rapidly emerging AI wearable audio market.

The AW88188 is built on a 55nm BCD (Bipolar-CMOS-DMOS) process, a mature but highly reliable manufacturing node ideal for mixed-signal audio applications. The chip boasts a static current of just 3.45mA, making it exceptionally power-efficient for battery-constrained wearable devices. Its integrated 200 MCPS (million cycles per second) digital signal processor enables sophisticated real-time audio processing, including AI-powered noise suppression and AI-driven sound field rendering for immersive spatial audio experiences.

Awinic's client roster reads like a who's who of the smart glasses industry. The company counts Meta, Rokid, Garmin, Xiaomi, XREAL, Lenovo, and RayNeo among its customers — collectively covering approximately 95% of the head-mounted display market. This near-total market penetration gives Awinic a formidable competitive moat as the AI glasses category expands.

"Awinic has been in the audio space for 18 years. We understand the acoustic challenges that wearable devices face — tiny speakers, limited battery, background noise, and the need for crystal-clear voice pickup," said a company representative at the forum. "The AW88188 is designed from the ground up for the AI glasses form factor."

Beyond its chip business, Awinic has also taken a strategic equity stake in Rokid, one of China's leading augmented reality glasses makers. This investment gives Awinic deeper insight into the design requirements of next-generation AI wearables and creates a vertically integrated relationship between chip designer and device manufacturer.

The AI glasses market is heating up globally, with Meta's Ray-Ban collaboration proving there is genuine consumer demand for smart eyewear. In China, companies like Rokid, XREAL, and Xiaomi are racing to deliver products that combine lightweight designs with always-on AI assistants. Audio quality — for voice commands, phone calls, music, and spatial notifications — has emerged as a critical differentiator.

Awinic's strategy is clear: become the indispensable audio chip supplier for this entire ecosystem. With 95% market coverage among head-mounted device makers and a chip purpose-built for the category, the company is well on its way to claiming the "Audio King" title it seeks.

The AW88188 is expected to begin sampling with key customers in the coming months, with mass production targeted for late 2026.

Tags: Awinic, AI Glasses, Audio Chip

DeepSeek's Coming of Age: Landmark Funding Round and the Case for Consumer Monetization

DeepSeek, the Chinese AI lab that has captured global attention with its cost-efficient large language models, is finally embracing outside capital in what could be one of the largest private fundraising rounds in AI history. The company is reportedly raising approximately 50 billion RMB (about $6.9 billion) at a valuation between 350 billion and 400 billion RMB ($48-55 billion), signaling a maturation for a firm that had long prided itself on being the last major AI company without external funding.

According to sources close to the matter, Tencent is evaluating an investment of roughly 10 billion RMB, while battery giant CATL is considering a 5 billion RMB stake. The round represents a pivotal moment for DeepSeek, which has operated with minimal outside capital, relying instead on the resources of its parent hedge fund, High-Flyer.

The funding comes at a critical juncture. DeepSeek has been grappling with severe compute shortages, leading to frequent service outages that frustrate its growing user base. Its latest model, DeepSeek V4, is widely regarded as one of the most cost-efficient AI models on the market, yet the company continues to slash prices rather than capitalize on its technological advantage.

Industry analysts argue this strategy is unsustainable. While competitor ByteDance has pushed forward with paid tiers for its Doubao assistant and other AI platforms have successfully monetized through subscription models, DeepSeek has remained conspicuously reluctant to charge its consumer users. The company's API pricing is among the most aggressive in the industry, but its consumer-facing service remains largely free, leaving significant revenue on the table.

"DeepSeek has built something truly remarkable in terms of cost-performance ratio, but they are leaving money on the table by not monetizing their C-end users," said one Beijing-based tech analyst. "Every other major AI assistant in China has introduced paid tiers. DeepSeek's reluctance to follow suit is becoming a competitive disadvantage."

The compute shortage issue further underscores the need for monetization. Without additional revenue streams, DeepSeek has struggled to secure the GPU capacity needed to serve its rapidly expanding user base during peak hours. The new funding round will help alleviate some of these infrastructure bottlenecks, but analysts argue that sustainable revenue generation through consumer subscriptions is essential for long-term viability.

As DeepSeek transitions from underdog to industry heavyweight, its ability to monetize its technological breakthroughs will determine whether it can compete with well-funded rivals like ByteDance, Baidu, and Alibaba. The funding round marks the end of DeepSeek's bootstrap era, but the real test lies ahead: can the company learn to charge for what it has given away for free?

Tags: DeepSeek, AI Funding, AI Monetization

Researchers at the University of Science and Technology of China (USTC) have open-sourced a novel agent-driven long-context training paradigm that achieves breakthrough efficiency — a 30-billion-parameter model matching the performance of Alibaba's Qwen3-235B, which is nearly eight times larger.

The core innovation lies in how the training data is sourced and structured. Traditional approaches to building long-context capabilities fall into two camps, both with significant drawbacks. The first is expensive manual labeling, where human annotators painstakingly craft long-context examples — a process that does not scale well. The second is heuristic short-text concatenation, which stitches together unrelated snippets but fails to produce the coherent, dependency-rich sequences that models need to learn genuine long-range reasoning.

The USTC team took a fundamentally different approach: they turned to AI agent trajectories. Instead of contriving long-context data artificially, they compiled the multi-turn interaction histories produced by autonomous agents as they navigate real tasks. These trajectories naturally contain the kind of extended, context-dependent exchanges that are precisely what long-context training requires — sequences of observations, reasoning steps, and actions that build on information introduced many turns earlier.

By treating agent trajectories as a first-class data source, the paradigm directly addresses what the researchers identify as the long-context capability bottleneck for AI agents. The resulting high-quality training data teaches models to maintain and manipulate information across extended contexts in a way that feels organic rather than manufactured.

The results speak for themselves. With only 30 billion parameters, the USTC-trained model achieves performance parity with Qwen3-235B on a range of long-context benchmarks. This represents a dramatic improvement in efficiency — a roughly 8× reduction in model size without sacrificing capability. For practitioners, this means that long-context reasoning that previously required enormous, compute-intensive models is now accessible with a far smaller footprint.

The open-source release allows the broader AI community to build on this work, potentially accelerating progress in agent-based systems, long-document understanding, multi-turn dialogue, and any application where maintaining coherence across extended interactions is critical. By demonstrating that data quality can substitute for raw scale, the USTC team has offered a compelling path forward for making capable long-context agents more widely accessible.

The paradigm shift is clear: rather than scaling models ever larger to handle longer contexts, we can train smaller models more intelligently — using the natural structure of agent behavior as our teacher.

Alibaba's flagship AI assistant, Qwen (also known as Tongyi Qianwen), has taken a major step toward becoming an open platform by opening up to third-party Agents and Skills. The move marks a significant shift in China's AI assistant landscape, positioning Qwen as a hub where enterprises can build, deploy, and operate branded AI agents directly within the app.

What's Changing

The Qwen app now allows enterprises to integrate their own Skills — purpose-built capabilities that users can invoke conversationally — and operate fully branded Agents on the platform. This means businesses can maintain their own identity and control over the user experience while leveraging Qwen's large user base and AI infrastructure.

Early Partners

The first wave of enterprise partners includes several household names across China's consumer economy:

• Luckin Coffee — users can order in-store pickup drinks directly through Qwen
• KFC (Yum China) — similarly, in-store pickup orders can be placed via the Qwen assistant
• China Eastern Airlines — exploring agent-based flight services and customer support
• MIXUE — the popular tea and ice cream chain is also joining the platform

These integrations go beyond simple chatbot interactions. Users can complete real transactions — ordering coffee, fried chicken, or bubble tea for pickup — all within the conversational interface of Qwen. It's a glimpse of a broader trend: AI assistants evolving from answer engines into action platforms.

The Bigger Picture

Qwen's opening represents a major opening of one of China's leading AI assistant platforms. By allowing third-party enterprises to deploy branded agents and custom skills, Alibaba is betting that Qwen can become a super-app-style gateway for AI-powered services. For enterprises, it offers a low-friction way to experiment with AI-driven customer interaction without building a standalone app from scratch.

More partners are in the pipeline, with several enterprises currently testing agent integration and planning to launch branded agents on Qwen soon. As the platform matures, we can expect a rapidly expanding ecosystem of skills — from travel booking and food delivery to financial services and beyond.

This is a clear signal that the battle for AI platform dominance in China is shifting from raw model capability to ecosystem breadth. Qwen is no longer just a chatbot — it's becoming a marketplace for AI-powered services.

For seven years running, the Zhiyuan Conference has stood as one of artificial intelligence's most academically uncompromising gatherings. Organized by the Beijing Academy of Artificial Intelligence (BAAI), this annual event has drawn participation from 14 Turing Award winners over its history — a track record that cements its reputation as a true intellectual checkpoint for the global AI community.

The 8th edition, taking place June 12–13, 2026, returns with a focused theme: brain-inspired intelligence and next-generation AI paths. It is not a conference for hype. Attendees and presenters alike describe it as "academically hardcore" — a place where rigorous science takes precedence over product launches and marketing narratives.

This year's program brings together more than 200 leading experts and over 40 AI company CEOs and co-founders, spanning the full spectrum of artificial intelligence research. The agenda cuts across five core frontier areas:

• Agent — autonomous systems capable of planning, reasoning, and executing complex tasks
• World Model — building internal representations that let AI understand and simulate reality
• Embodied AI — intelligence that interacts with the physical world through robotics and sensors
• AI Self-Evolution — systems that improve their own capabilities through continuous learning
• AI Safety — alignment, robustness, and the governance of increasingly capable models

Each of these tracks reflects a deliberate shift away from scaling alone. While large language models have dominated headlines in recent years, the Zhiyuan Conference has consistently pushed the conversation toward architectures that draw inspiration from biological intelligence — neural circuits, cognitive architectures, and learning mechanisms that go beyond next-token prediction.

BAAI, the organizing body behind the conference, has positioned itself as a leading force in open-source AI research and open science. The Zhiyuan Conference serves as its flagship forum for surfacing ideas that are still too early, too ambitious, or too foundational for mainstream industry conferences. It is where researchers share raw results, debate architectural directions, and test the limits of current paradigms.

For AI insiders, the conference functions as an annual barometer — a place to measure how far the field has come and, more importantly, to sense where it is heading next. The 2026 edition promises to be no exception. As the community grapples with questions of safety, scalability, and the nature of intelligence itself, the conversations in Beijing this June will likely echo through the rest of the decade.

Announced at NVIDIA GTC Taipei 2026, the achievement marks a significant milestone for China's embodied intelligence sector.

TAIPEI — A Chinese embodied AI company has claimed the top spot on RoboArena, widely regarded as the world's hardest-to-cheat benchmark for embodied artificial intelligence, outperforming industry giants NVIDIA and Physical Intelligence (PI). The announcement was made during NVIDIA's GTC Taipei 2026 conference.

RoboArena is designed to test embodied AI systems across diverse real-world physical tasks, evaluating not just perception and planning but actual robotic execution in unstructured environments. Its rigorous methodology — randomized environmental conditions, adversarial perturbations, and strict reproducibility — has earned it a reputation as the benchmark least susceptible to overfitting.

The Chinese company demonstrated superior performance across multiple task categories including object manipulation, mobile navigation, and complex tool use. Their solution leverages a novel architecture that tightly integrates visual perception with motor control at the neural network level, enabling faster adaptation to novel environments with minimal fine-tuning.

NVIDIA's entry, built on its Isaac robotics platform and GPU-accelerated simulation-to-real pipelines, was a strong contender but ultimately fell short of the top aggregate score. Physical Intelligence, known for its work on generalist robot policies, also placed behind the Chinese contender.

The result challenges the assumption that massive compute budgets are the primary driver of embodied AI performance. The Chinese startup's victory suggests architectural innovation and efficient real-world data utilization can overcome significant resource disadvantages. As one industry analyst noted, "This demonstrates that the field remains wide open."

For the winning Chinese company, the RoboArena #1 ranking provides powerful third-party validation and is likely to accelerate investor interest. With the embodied AI market projected to grow exponentially, this David-versus-Goliath story highlights China's rapidly maturing embodied AI ecosystem, combining world-class hardware manufacturing, deep computer vision talent, and aggressive government support for AI industrialization.

On June 2, Shanghai-headquartered UCloud cut the ribbon on a new cloud computing node in Uzbekistan, pushing its global network to 36 nodes across 28 regions. The move might once have been a routine infrastructure announcement. In 2025, it is something larger: a signal that the logic of Chinese cloud computing exports has been fundamentally rewritten by the large language model explosion.

UCloud, one of China's earliest independent cloud providers, built its reputation serving internet and gaming companies scaling across Southeast Asia and the Middle East. Its new Central Asian node — covering Uzbekistan, Kazakhstan, and neighboring markets — gives local enterprises and Chinese companies expanding along the Belt and Road low-latency access to GPU-powered compute.

"Previously, competition was about slicing up existing demand," a UCloud executive noted. "AI brings massive incremental growth."

That shift is visible across Shanghai's cloud sector. Where Chinese cloud providers once exported to serve overseas Chinese enterprises and local e-commerce, the calculus has changed. The primary export today is AI infrastructure itself — GPU clusters, high-performance networking, and the cooling and power systems needed to train and serve frontier models. Shanghai, home to a dense concentration of data center know-how and chip design talent, has become a funnel for compute capacity flowing into markets that lack domestic alternatives at scale.

UCloud's Uzbekistan node offers bare-metal GPU instances and managed Kubernetes optimized for inference workloads. The company's strategy mirrors broader trends: Alibaba Cloud and Tencent Cloud are both investing heavily in Southeast Asian and Middle Eastern AI regions, and smaller providers like UCloud are carving niches in underserved frontiers — Central Asia, Africa, Latin America — where hyperscalers have been slow to plant flags.

The stakes go beyond any single node. Chinese cloud exports are no longer about undercutting AWS on price for basic storage and virtual machines. They are about exporting the entire stack — the hardware, the orchestration, the model-hosting platforms — that powers the AI economy. For Shanghai's computing industry, the Uzbekistan launch is less a geographic expansion than a proof point: the cloud export playbook has been rewritten, and the new chapter runs on GPUs.