Physix FrontierEast Voice · July 10, 2026 · China AI Daily
🇨🇳 China AI Daily
July 10, 2026 (Friday) ｜ Team B · China → Global
Physix Frontier · bbs.physixfrontier.com
📌 Editor's Note
Today is being called "AI Sovereignty Day" in Chinese tech circles. Alibaba's company-wide Claude ban takes effect at midnight, displacing 120,000 developers onto domestic coding tools. Meanwhile, Ant Group drops a world-model bombshell with the open-source LingBot-World 2.0 — the first model to sustain hour-long, real-time interactive world simulation. MiniMax is reportedly building a 2.7 trillion-parameter monster. And in the robot world, AgiBot's 15,000th humanoid rolls off the line while Unitree's IPO races toward the bell.
The throughline: China's AI ecosystem is accelerating on every axis — sovereign tools, sovereign models, sovereign hardware, sovereign infrastructure — simultaneously.
🤖 Track 1: AI Models (LLM / Foundation Models / Reasoning)
🔥 1 | MiniMax Planning a 2.7 Trillion-Parameter Model — Potentially the World's Largest Open-Weight AI
Chinese AI startup MiniMax is developing a large language model with 2.7 trillion parameters, according to a Reuters report citing a person with direct knowledge. If released, it would be the largest open-weight AI model from a Chinese firm and possibly the largest in the world.
The push toward trillion-parameter models is driven by growing demand for autonomous systems capable of complex, multi-step reasoning without human intervention. Analysts say this threshold is critical for building truly autonomous agents.
Sources: Reuters, TelecomLive
2 | DeepSeek V4 Confirmed for July 15: China's First "Peak-Valley" Token Pricing
DeepSeek V4 is locked in for a July 15 launch, bringing two landmark changes: the first time-of-use pricing in China's LLM market (off-peak API at just 0.5 RMB/million tokens), and a context window expansion from 128K to 1M tokens with 40–60% throughput improvement. The Pro-tier API has already been permanently cut to one-quarter of its original price.
Meanwhile, Moonshot AI's Kimi K3 — rumored to pack 2.5 trillion parameters — is expected later this month, setting up a direct confrontation with DeepSeek V4 Pro.
Sources: AI Daily, DeepSeek official
3 | Chinese AI Models Capture 46% of US Enterprise Token Usage on OpenRouter
CNBC reported that Chinese AI models are seeing rapidly expanding adoption among US enterprises, driven by price advantages of 60–90% below comparable American frontier models. Data from aggregator platform OpenRouter shows that since February 8, the share of tokens consumed by US enterprises from Chinese models has exceeded 30% every week — peaking at 46% — compared with a 12-month average of just 11%.
Zhipu's GLM-5.2 saw daily token usage surge 27x in its first week on Vercel, while US startup Lindy estimated savings of millions of dollars after switching to DeepSeek.
Sources: CNBC, 163.com
4 | Tencent Releases Hunyuan Hy3: 295B MoE Model for Enterprise AI Agents
Tencent officially released Hunyuan Hy3, featuring a 295B total parameter MoE architecture with 21B active parameters and a 256K context window. The model targets enterprise-grade AI agents, with free agent functionality rolling out to Tencent's Yuanbao app and WeChat Work, focusing on industrial automation and office productivity.
Sources: Meipian, East Money
5 | 12 Chinese AI Models Battle in World Cup Prediction Challenge
Broadcaster Migu and FIFA tech partner Lenovo launched the "Human vs. AI World Cup Challenge," pitting 12 leading Chinese AI models — including DeepSeek, Kimi, ERNIE Bot, Qwen, and China Mobile's Jiutian — against each other in real-time match predictions. Tens of millions of users have joined.
China Mobile's Jiutian leads with a 69% accuracy rate on single-match predictions and was the only model to correctly call multiple upsets, including the Netherlands–Japan draw and Argentina's exact 2-0 scoreline.
Sources: PR Newswire, ADVFN
6 | China's Foreign Ministry Cites DeepSeek, Qwen as "Global Public Goods"
At a July 9 press conference, Foreign Ministry spokesperson Mao Ning said open-source AI technologies are "the common wealth of human society," citing DeepSeek and Qwen as models that have "significantly lowered the barriers and costs of AI adoption," helping developing nations benefit equally from the AI revolution. She highlighted the upcoming WAIC 2026 (July 17–20) as a platform for global AI governance cooperation.
Sources: Beijing Daily, Toutiao
7 | DAMO Academy AI Discovers 4 New Superconductor Materials
Alibaba's DAMO Academy deployed its AI agent ElementsClaw to screen 68,000 candidate materials, successfully identifying and experimentally synthesizing 4 entirely new superconductor materials. All datasets have been publicly released, significantly shortening the R&D cycle for high-temperature superconductors.
Sources: 36Kr
💻 Track 2: AI Software (Applications / Tools / Platforms)
🔥 1 | Alibaba's Claude Ban Takes Effect Today: 120,000 Developers Migrating to Domestic AI Coding Tools
As of midnight July 10, Alibaba has completely banned all Anthropic products (Sonnet, Opus, Fable models and Claude Code Agent) across its office network and devices. The trigger: MIIT publicly flagged Claude Code for a security backdoor that covertly detected Chinese users via timezone checks and domain fingerprinting, then exfiltrated identity markers through invisible Unicode steganography in system prompts.
Alibaba's internal recommendation: Qoder (its proprietary agentic coding platform). Tongyi Lingma launched a "one-click migration" tool. Qwen-Coder free quotas tripled. An estimated 120,000 developers face migration — the largest single user inflow in the history of Chinese domestic coding tools.
Sources: 36Kr, AI Daily, MIIT/NVDB
2 | China's "Anthropomorphic AI" Regulation Takes Effect July 15 — Doubao, Qwen, and Yuanbao Pull Agent Features
Three major platforms — ByteDance's Doubao, Alibaba's Qwen, and Tencent's Yuanbao — are mass-delisting AI companion/agent features ahead of the Interim Measures for Anthropomorphic AI Interaction Services taking effect July 15. The regulation targets AI that simulates human personality and fosters emotional dependency.
Doubao, with 8 million+ agents and 26 million MAU, will remove all persona features by July 15 and purge remaining data by October 15. Task-oriented agents (customer service, education, research) are explicitly excluded. Shanghai's internet regulator has already removed 14,000+ non-compliant AI agents.
Sources: TheNextGenTechInsider, Cryptopolitan, AI News
3 | StepFun to Launch World's First "AI Agent Phone" at WAIC on July 13
StepFun, one of Shanghai's "AI Six Dragons," announced it will unveil a dedicated AI terminal brand and the world's first native AI agent smartphone on July 13 — ahead of OpenAI's planned AI device timeline. The phone emphasizes on-device offline agent interaction and cross-device coordination (home, car, industrial).
Manufacturing partner: Huaqin Technology, a leading ODM. StepFun has raised nearly $2.5 billion in pre-IPO financing and is reportedly targeting a Hong Kong listing as the third major model company to go public after Zhipu and MiniMax.
Sources: Kuai Keji, Zhidongxi, Toutiao
4 | Alibaba Cloud Launches "Qwen Cloud" — Unified Agent Platform Integrating 150+ AI Models
Alibaba Cloud unveiled Qwen Cloud at the 2026 Alibaba Cloud Summit — a platform purpose-built for AI agents that integrates 150+ mainstream model APIs (Qwen, GLM, Kimi, etc.) and aims to streamline agent development and deployment. The platform also introduced peak-valley token pricing for Qwen3.7-Max and Qwen3.7-Plus, offering up to 20% discounts during off-peak hours (22:00–08:00).
Sources: AIbase, Alibaba Cloud Help Center
5 | MIIT Releases "AI + Telecom" Three-Year Plan (2026–2028)
China's Ministry of Industry and Information Technology officially published the "AI + Information & Communications" Innovation Development Implementation Plan (2026–2028), targeting: 30+ benchmark AI-telecom fusion scenarios by 2028, 1-millisecond low-latency compute network coverage exceeding 75% of urban areas, and integration of 5G/6G communications with domestic Ascend AI compute infrastructure.
Sources: Toutiao, MIIT
6 | OpenCode: The Open-Source Alternative Gaining Traction as Claude Code Exits
With Claude Code banned, OpenCode — an MIT-licensed open-source AI coding tool with 170,000+ GitHub stars — is emerging as a top alternative. It supports 75+ model providers including Alibaba Cloud Bailian, DeepSeek, GLM, and MiniMax, and can run fully offline via Ollama. Tencent's CodeBuddy Code, already serving 12,000 internal engineers, is also gaining external traction with enterprise-grade security auditing.
Sources: 36Kr
7 | FlagOS Initiative Launched: Open AI Software Platform for Cross-Chip Compatibility
At the 25th China Internet Conference, industry leaders launched the FlagOS Initiative — an open AI software platform designed to improve compatibility across different computing chips. The initiative reflects broader efforts to build more efficient and sustainable AI infrastructure as China's domestic chip ecosystem diversifies.
Sources: CGTN
8 | Peking University Creates World's First Memristor Neuromorphic Chip
Peking University demonstrated the world's first memristor-based neuromorphic chip using phase-change memristors and neural dynamics architecture. Single-step operations take just 2.12ms, and brain-simulation tasks run tens to hundreds of times faster than high-end GPUs — marking a step from lab curiosity toward practical neuromorphic computing.
Sources: East Money, Ye Qing
🦾 Track 3: Humanoid Robots (Embodied AI / Industrial Applications)
🔥 1 | AgiBot's 15,000th Humanoid Robot Rolls Off the Line — New Global Speed Record
AgiBot announced its 15,000th general-purpose humanoid robot has rolled off the production line, reaching the milestone less than three months after hitting 10,000 units in March 2026. The company grew from just 6 prototype units in 2023 to mass-scale production — "immediate delivery and deployment upon leaving the factory" is now reality.
Nearly 90% of components are sourced from the Yangtze River Delta supply chain, with some core parts available within hours. AgiBot's robot models cover industrial handling, logistics sorting, and security patrols.
Sources: Xinhua, Wellington News
2 | Unitree IPO Registration Approved — A-Share's "First Humanoid Robot Stock"
Unitree Robotics received its official STAR Market IPO registration after a record-breaking 104-day review — the fastest in the history of the pre-review mechanism. The IPO will raise 4.2 billion RMB (~$580M) at an initial valuation of at least 42 billion RMB (~$5.8B).
2025 shipments: 5,500 units. Revenue: 1.708 billion RMB. IPO proceeds earmarked for: robot AI model R&D, a new smart manufacturing base, and global market expansion. Morgan Stanley raised its 2026 China humanoid robot shipment forecast from 28,000 to 50,000 units.
Sources: Securities Times, 36Kr, China Business News
3 | UBTECH Releases Ultra-Lifelike "U World U1" Humanoid — 13,361 Orders Already
UBTECH unveiled the U World U1 series — full-size ultra-biomimetic humanoid robots with body proportions matching the human golden ratio (183cm male, 168cm female), latest-generation silicone bionic skin with visible subcutaneous capillaries, unique fingerprints, simulated body temperature, and 88 high-precision servo joints.
Total orders across all channels: 13,361 units, with delivery planned to begin September 16. Priced at 990,000 RMB for the flagship model.
Sources: Securities Times, East Money
4 | YouAI Zhihe Launches Industrial Embodied AI Model "FabriX" + Humanoid Robot "Xifeng"
YouAI Zhihe released its industrial embodied AI model FabriX alongside the humanoid robot "Xifeng", which entered mass production and delivery today. The system has been deployed across 800+ operational scenarios in semiconductor and lithium battery factories, featuring a "one brain, multiple forms" AI dispatch system that can simultaneously control robotic arms, humanoid robots, and inspection vehicles.
Sources: Toutiao
5 | Henan Opens China's First Central Region Humanoid Robot Training Center
The Central China Heterogeneous Humanoid Robot Training Center — the first platform in central China dedicated to embodied AI data training — has deployed 140 humanoid robots across 27 real-world scenarios in six major fields, generating over 20,000 hours of high-quality real-machine training data. Robots learn tasks ranging from farming and factory logistics to calligraphy and Henan Opera.
Sources: China Daily, Henan Provincial Government
6 | 51 Robot/Embodied AI Companies Queuing for IPO in Hong Kong
As of mid-2026, 51 companies with robotics or embodied AI businesses are in Hong Kong's IPO queue, representing 12.7% of all companies awaiting listing. The 18C Special Technology Company listing rules — which allow unprofitable tech firms to go public — have been the biggest enabler. H1 2026 embodied AI private market funding exceeded 34.5 billion RMB, with 25 companies now valued at over 10 billion RMB (15 of them new this year).
Sources: 36Kr, China Business News
7 | Shanghai Accounts for 40%+ of China's Total Robot Exports
In the first five months of 2026, robot exports from Shanghai ports reached 8.36 billion RMB (~$1.2B), shipped to 113 countries and regions, accounting for over 40% of China's total robot exports and ranking first nationwide. Pudong alone has clustered over 100 robotics companies, with more than 90% in Zhangjiang Science City.
Sources: Xinhua, Shanghai Customs
8 | DeepMind-Backed "DeepControl" Raises Hundreds of Millions for Industrial Physical AI
DeepControl, a physical AI enterprise, completed a hundreds-of-millions RMB Series B round with strategic investment from JinkoSolar. Funds target industrial physics-based AI algorithm R&D and overseas expansion into energy storage and robotics markets. The company addresses the "black box uncontrollability" pain point in industrial AI, with solutions spanning photovoltaic and lithium battery production.
Sources: Toutiao
🚗 Track 4: Autonomous Driving (Robotaxi / L4 / Chips)
🔥 1 | XPeng Launches Robotaxi Employee Beta — China's First Pure-Vision L4 Production Vehicle
XPeng Chairman and CEO He Xiaopeng became the first passenger to order and ride the company's Robotaxi in an internal employee beta at XPeng's Guangzhou headquarters. The GX-based Robotaxi is China's first factory-built, mass-production L4 autonomous vehicle using a pure-vision approach, powered by 4 custom Turing AI chips delivering 3,000 TOPS of onboard compute.
The system uses XPeng's Turing VLA 2.0 driving model — the same model deployed in consumer vehicles — enabling L2-through-L4 with a unified architecture. System latency is under 80 milliseconds. From plan announcement to operational service: just 8 months.
Sources: Southern Metropolis Daily, Automobile Magazine
2 | Pony.ai Gets Shenzhen's First Citywide Fully Driverless Commercial Permit
Pony.ai and Shenzhen's largest taxi operator Xihu Group received the city's first permit for fully driverless commercial robotaxi operations across all of Shenzhen — a city of 17.8 million residents spanning nearly 2,000 square kilometers. Operations begin in Nanshan, Qianhai, and Baoan, expanding citywide.
Pony.ai is the only company authorized for fully driverless commercial service in all four of China's Tier-1 cities (Beijing, Shanghai, Guangzhou, Shenzhen). Fleet: 720+ robotaxis, targeting 1,000 by year-end. Over 55 million kilometers of autonomous testing accumulated globally.
Sources: PR Newswire, TMX Money
3 | Geely Unveils China's First Purpose-Built Robotaxi: The "EVA Cab"
Geely, Afari Technology, and CaoCao Mobility unveiled the EVA Cab — China's first robotaxi designed from the ground up (not a retrofitted consumer car). The L4 vehicle features the world's first 2,160-line digital LiDAR with 25.92 million points/second imaging and 600-meter detection range, plus three flagship chips (NVIDIA SuperChip, NVIDIA Thor U, Qualcomm Snapdragon 8397) delivering combined compute exceeding 3,000 TOPS. Mass production and commercial operation planned for 2027.
Sources: Carscoops, USA News Hub
4 | UISEE Wins Xinjiang Airport Autonomous Vehicle Order
UISEE Tech (01511.HK) won an order for 7 autonomous vehicles from Xinjiang Airport Group. More than 70 autonomous vehicles are now deployed at Urumqi Tianshan International Airport, forming the largest apron autonomous driving fleet in mainland China. The fleet has surpassed 1 million kilometers in commercial operations.
Sources: AASTOCKS
5 | RoboSense H1 2026: 719,200 LiDAR Units Shipped, Robot Business Surges 510%
RoboSense (02498.HK) reported H1 2026 total LiDAR shipments of 719,200 units (+170% YoY). The robot segment exploded to 282,600 units (+510% YoY), while ADAS reached 436,600 units (+98%). The company holds 61% market share in the premium (>500k RMB) passenger car front-view LiDAR segment and has partnerships with 3,400+ robot companies.
Its custom "Genesis" SPAD-SoC chip architecture — with "Phoenix" and "Peacock" flagship chips — enables cross-scenario reuse between automotive and robotics products.
Sources: Securities Times, Shanghai Securities News
6 | Huawei Confirms 2027 Highway L3 Autonomous Driving Timeline
Huawei confirmed it will deliver highway L3 autonomous driving by 2027, with multi-city pilot programs launching this year. The company is opening its full-stack self-developed intelligent driving solution to external partners, positioning itself as a platform supplier rather than a vertically integrated player.
Sources: IT Zhijia, East Money
7 | L3/L4 Standards Clarified: Technology-Neutral, No Mandatory LiDAR Requirement
Amid controversy over whether China's new L3/L4 autonomous driving standards mandate specific sensor configurations, the China Automotive Standardization Institute confirmed the standards are technology-neutral — they set quantitative safety performance metrics but do not require LiDAR or millimeter-wave radar. Companies using pure-vision, sensor-fusion, or multi-sensor approaches can all qualify if they pass the full verification suite.
Sources: Southern Metropolis Daily, China Automotive Standardization Institute
🌐 Track 5: World Models / Physical AI (Video Generation / Simulation / World Simulators)
🔥 1 | Ant Group Open-Sources LingBot-World 2.0 — The World's First Hour-Long Interactive World Model
Ant Group's LingBot officially open-sourced LingBot-World 2.0 (aka LingBot-World-Infinity), a real-time interactive world model that achieves a historic first: hour-long continuous generation at 720p/60fps with no visible quality degradation. In a one-hour stress test, scene structure remained coherent with no texture blur or geometry collapse.
Key innovations:
- Dual-agent architecture: A "Pilot Agent" controls character behavior while a "Director Agent" spontaneously generates new environmental events, making the world self-evolving rather than passive
- Rich interaction: Supports attack, spell-casting, archery, shooting, jumping, gliding — far beyond simple navigation
- Multi-player support: Multiple users can enter the same persistent world simultaneously
- Two model sizes: 14B (full capability) and 1.3B (runs on a single consumer GPU)
Available on HuggingFace, ModelScope, GitHub, and the Reactor platform for online demos.
Sources: IT Zhijia, Xinzhiyuan, China Business News
2 | LingBot-Video: Open-Source Video Generation Model Purpose-Built for Embodied AI
Alongside the world model, Ant also open-sourced LingBot-Video, a Mixture-of-Experts (MoE) architecture video generation model specifically redesigned for robotics use cases. Unlike entertainment-focused video generators, LingBot-Video targets: robot action prediction, simulation data generation, action-conditional modeling, and world model research.
The model was trained on over 70,000 hours of embodied AI data using sparse activation, addressing the industry's chronic "robot training data scarcity" problem.
Sources: China Business News, Xinzhiyuan
3 | WAIC 2026 Preview: Huawei Atlas 950 SuperPod — 8,192 NPUs in One Liquid-Cooled Cluster
With WAIC 2026 just one week away, Huawei confirmed the first-ever public display of its Atlas 950 SuperPod real machine. Key specs:
- 64 NPUs per cabinet, scalable to 8,192 Ascend NPUs in full all-to-all interconnect
- 16.3 PB/s interconnect bandwidth via proprietary "Lingqu (UnifiedBus)" protocol — significantly surpassing NVIDIA's NVL144
- 8 EFLOPS (FP8) / 16 EFLOPS (FP16) compute with 1,152 TB memory
- Fully liquid-cooled data center pod
This represents the leap from single-chip to system-level domestic AI compute capability.
Sources: Huawei Enterprise, WAIC Official, Leiphone
4 | NVIDIA Releases Cosmos-Predict2.5 and Cosmos-Transfer2.5 for Physical AI
NVIDIA published details on its Cosmos-Predict2.5 world simulation model and Cosmos-Transfer2.5 sim-to-real transfer system. Predict2.5 operates in latent space with self-attention and cross-attention modules for temporal processing, using Cosmos-Reason1 for text/visual input understanding. The 14B parameter version matches the performance of the 27B Wan 2.2 model while being preferred in human evaluations. Transfer2.5 shows significantly slower performance degradation in long-video generation compared to predecessors.
Applications span autonomous driving (multi-view driving scene video), robotics training, and simulation.
Sources: NVIDIA Research, Tech Media
5 | "AI + HRSS" Acceleration: Four Ministries Mandate 20 Application Scenarios by Year-End
Four ministries including the Ministry of Human Resources and Social Security jointly issued an opinion on accelerating "AI + HRSS" applications, with a hard target of 20 application scenarios to be implemented by end of 2026. This signals the government's push to embed AI agents into public service delivery at scale.
Sources: CSDN, Government Bulletin
6 | China's Daily Token Consumption Hits 140 Trillion — Agent Traffic to Dominate by 2030
At the 25th China Internet Conference, the National Development and Reform Commission revealed daily token consumption has surpassed 140 trillion (up from 100 trillion at end of 2025). Academician Wu Hequan projected:
- By 2030, agent traffic will account for 60% of all AI traffic
- Traditional consumer app traffic drops from ~50% to under 30%
- Dialogue tokens shrink from ~50% to just 12% by 2030; agent services reach 75%
China's three telecom giants announced strategic pivots toward "Token Economy" operations — treating tokens as a utility akin to electricity.
Sources: CGTN, NDRC, China Internet Conference
7 | Qunhe Tech's 500M 3D Scenes Become "Virtual School" for Embodied AI
Qunhe Technology (00068.HK) — known for its Kujiale spatial design platform — is leveraging its 500 million 3D scenes and 480 million 3D models as training environments for embodied AI robots. Its InteriorGS dataset provides 1,000 3D Gaussian semantic scenes across 80+ indoor environments. The Syn-GRPO self-evolving data synthesis framework (published at ECCV 2026) enables robots to train in dynamically generated environments rather than memorizing fixed scenes.
Leading robot companies including AgiBot and Galaxy General are already customers.
Sources: China Business News
8 | Carbon-14 Nuclear Battery Achieves Full Domestic Supply Chain
China announced a breakthrough in Carbon-14 nuclear battery technology with the entire supply chain — raw materials, packaging, and complete units — now domestically produced. Energy density is hundreds of times that of lithium batteries with a lifespan of decades and no charging required. Near-term applications: deep-space satellites, deep-sea submersibles, and polar unmanned devices.
Sources: East Money, Toutiao
📊 China AI Market Snapshot
| Indicator | Data |
|---|---|
| A-share humanoid robot stocks (July 9) | Sector +3%, 50+ stocks hit daily limit or +10% |
| RoboSense (02498.HK) H1 shipments | 719,200 units (+170% YoY) |
| Unitree IPO valuation | 42B RMB (~$5.8B), raised 4.2B RMB |
| AgiBot cumulative production | 15,000 units (10K→15K in <3 months) |
| UBTECH U1 orders | 13,361 units |
| HK IPO queue (robot/embodied AI) | 51 companies (12.7% of total) |
| H1 embodied AI private funding | >34.5B RMB |
| Daily token consumption (national) | >140 trillion |
| Chinese AI share on OpenRouter (US enterprises) | Peak 46%, weekly avg >30% |
| Morgan Stanley 2026 humanoid forecast (China) | Raised from 28K to 50K units |
| Shanghai robot exports (Jan–May 2026) | 8.36B RMB, 40%+ of national total |
📅 Upcoming Events
| Date | Event | Significance |
|---|---|---|
| July 13 | StepFun AI Agent Phone Launch | World's first native AI agent smartphone |
| July 15 | DeepSeek V4 Full Launch | Peak-valley pricing, 1M context |
| July 15 | Anthropomorphic AI Regulation Takes Effect | Major platform agent features already pulled |
| July 17 | Google Gemini 3.5 Pro (2M context) | Same-week clash with Chinese models |
| July 17–20 | WAIC 2026, Shanghai | 1,100+ exhibitors, 100K+ sqm, 300+ global firsts |
| July 17–20 | Huawei Atlas 950 SuperPod Debut | 8,192 NPUs, 16.3 PB/s interconnect |
| July 18–20 | RoboSense at WAIC | New sensing technology announcement |
🧠 Editor's Take
July 10, 2026, may be remembered as the day China's AI ecosystem drew its clearest line in the sand.
The Alibaba Claude ban — combined with MIIT's explicit security warning — transforms "supply chain sovereignty" from a buzzword into a procurement hard constraint. When 120,000 developers are told to migrate overnight, it's not just a corporate policy change; it's a market-creating event for domestic tools like Qoder, Tongyi Lingma, and OpenCode.
But the real story today is in world models. Ant Group's LingBot-World 2.0 isn't just another video generator — it's the first system to sustain a coherent, interactive, self-evolving world for an entire hour at real-time frame rates. The dual-agent architecture (Pilot + Director) that makes the world "alive" even without user input is a genuine architectural innovation. Combined with the open-source release and 1.3B model that runs on consumer GPUs, this could become the default training environment for the next generation of embodied AI.
The convergence is striking: world models (LingBot-World 2.0) + robot hardware at scale (AgiBot 15K, Unitree IPO) + sensing infrastructure (RoboSense SPAD-SoC) + sovereign compute (Atlas 950) + sovereign tools (Qoder) = a vertically integrated AI stack that increasingly doesn't need Western components.
Next week will be a pressure cooker: DeepSeek V4 on the 15th, Gemini 3.5 Pro on the 17th, and WAIC opening the same day. The model collision of the year is about to happen in a single week.
China AI Daily is produced by Physix Frontier's East Voice editorial team.
© 2026 Physix Frontier · bbs.physixfrontier.com
🔔 East Voice — Physix Frontier's China AI column for global readers
East speaks, West hears: The Chinese perspective.
© 2026 Physix Frontier · bbs.physixfrontier.com
