Physix FrontierEast Voice · China AI Daily · August 2, 2026
East Voice — China AI Daily Digest
August 2, 2026 (Sunday)
Editor's Note
DeepSeek V4-Flash officially launches with dramatically enhanced Agent capabilities; Tencent Hunyuan Hyra solves a 50-year-old open math problem; ByteDance releases Seedance 2.5 video model; SenseTime DaXiao open-sources embodied foundation model ACE-Brain-0.5, outperforming GPT-5.4 and Gemini at only 8B parameters; BYD's humanoid robot "Xiaodi" to debut in August; Magic Atom HQ breaks ground in Wuxi with 10,000-unit annual capacity.
Editor's Take: China's AI industry is shifting from a "parameter arms race" to "scenario validation." DeepSeek's Agent capabilities, SenseTime's 8B embodied model, and JD.com's 10M+ device integration — competing on real-world impact rather than benchmarks is the hallmark of a maturing industry.
I. AI Foundation Models (LLM)
1. DeepSeek V4-Flash Official API Launches with Dramatically Enhanced Agent Capabilities
Date: 2026-07-31
Summary: DeepSeek V4-Flash official API goes live — 284B total parameters with only 13B active for inference. Agent capabilities surge: Terminal Bench 2.1 scores 82.7, Cybergym 76.7, Toolathlon Verified 70.3, full-stack dev benchmark DSBench-FullStack reaches 68.7. AI begins working like an engineer: opening terminals, analyzing repos, invoking tools, completing end-to-end tasks.
Source: Guangzhou Daily / Toutiao
Editor's Take: 13B active parameters scoring 82.7 on Terminal Bench — MoE architecture's efficiency advantage is fully validated in Agent scenarios.
2. Tencent Hunyuan Hyra Solves 50-Year-Old Open Problem in Additive Combinatorics
Date: 2026-07-30
Summary: Tencent Chief AI Scientist Yao Shunyu reveals that Hunyuan research agent Hyra (built on Hy3) has solved a ~50-year-old open problem in additive combinatorics, constructing integer set families that prove optimal exponents. Among 55 long-standing open math problems, Hyra set new best-known results on 29. Paper posted on arXiv with Lean 4 formal verification underway.
Source: Huxiu
Editor's Take: From "assisting research" to "independently solving open problems" — a milestone achievement for AI for Science. Hy3's call volume is 68x that of Hy2, topping OpenRouter globally.
3. ByteDance Seed Team Releases Seedance 2.5 Video Generation Model
Date: 2026-07-31
Summary: ByteDance releases next-gen video creation model Seedance 2.5, now available across Jimeng, Xiaoyunque, Doubao Pro, and Coze. Generates up to 30-second video clips in a single pass, supports up to 30 images + 10 video clips + 10 audio clips as reference materials, with precise timestamp-level control.
Source: Tencent News
Editor's Take: 30-second single generation with multi-modal input references — Seedance 2.5 takes a major step toward practical video generation.
4. MiniMax Launches H3 Omni-Modal Generation Model
Date: 2026-07-31
Summary: MiniMax releases H3, a universal omni-modal generation model that jointly understands text, image, video, and audio, generating native 2K resolution, 15-second video with native stereo audio. Open weights planned for near-term release. Targets advertising, e-commerce, and gaming scenarios.
Source: MiniMax official website
Editor's Take: Native stereo audio + 2K video in a single generation pass — strong commercial potential in advertising and e-commerce.
5. Tsinghua Open-Source Spatial Intelligence Model Accepted at ECCV 2026, Surpasses Gemini
Date: 2026-08-02
Summary: Tsinghua's open-source spatial intelligence model accepted at ECCV 2026, outperforming Gemini across all spatial understanding tasks on 120-minute long videos. Core breakthrough: spatial memory module + dynamic correction mechanism solving the "long-term spatial drift" problem. All code and weights open-sourced.
Source: Toutiao
Editor's Take: From "understanding images" to "understanding the spatial world across entire videos" — the 120-minute long-memory breakthrough is immediately usable infrastructure for robotics and autonomous driving developers.
6. SenseTime DaXiao Open-Sources Embodied Foundation Model ACE-Brain-0.5
Date: 2026-08-01
Summary: SenseTime subsidiary DaXiao Robotics open-sources ACE-Brain-0.5, a unified embodied foundation model. At only 8B parameters, it systematically outperforms GPT-5.4, Gemini-2.5-Pro, Claude-Sonnet-4.6, NVIDIA GR00T N1.6, and Physical Intelligence π₀/π₀.₅ across multiple international embodied intelligence benchmarks. Advances robot foundation models from "understanding the world" to an integrated "understand-plan-act-evaluate" closed loop.
Source: Toutiao
Editor's Take: 8B parameters beating 100B+ models — embodied intelligence isn't a parameter-counting game, architecture design matters more. SenseTime stock up 7% on the day.
7. CXMT LPDDR6 Memory Enters Final Validation Phase, Mass Production in H2
Date: 2026-08-02
Summary: China's next-gen low-power memory LPDDR6 enters final validation, with speeds up to 12,800 Mbps. Expected to enter mass production in H2 2026, enabling flagship smartphones and edge AI devices.
Source: Toutiao
Editor's Take: Domestic LPDDR6 mass production will lower the hardware threshold for on-device AI inference — significant for edge AI deployment.
II. AI Software (AI Application / SaaS)
8. Zhichi Tech Launches AI Agent-Driven Next-Gen Customer Contact Platform
Date: 2026-07-30
Summary: Zhichi Tech upgrades its product architecture to three pillars — Agents, Nexus, Experts — transforming from "unified customer contact SaaS" to "AI Agent-driven next-gen customer contact platform." Agents can independently handle customer interactions and proactively reach out, with Nexus providing a unified foundation across all channels.
Source: Toutiao
Editor's Take: The customer service paradigm shift from "humans + systems" to "agents + experts" is one of the most certain AI commercialization scenarios.
9. ByteDance Integrates Feishu and Doubao into "Creativity Services Platform"
Date: 2026-07-31
Summary: ByteDance merges Feishu (Lark) and Doubao into a "Creativity Services Platform" commercial organization, managed by Volcano Engine head Tan Dai. Over 90% of new Feishu enterprise customers in Q2 2026 simultaneously purchased AI office products.
Source: Multiple sources
Editor's Take: The Feishu+Doubao integration is a major To B event. 90% of new customers buying AI products simultaneously signals that "AI-native office" has shifted from add-on to must-have.
10. JD.com Goes "All In on Embodied Intelligence," JoyInside to Reach 10M+ Devices by Year-End
Date: 2026-08-01
Summary: JD.com founder Liu Qiangdong personally heads the Exploration Research Institute. JD launches "JoyInside" — modular AI capabilities embedded into existing smart hardware, partnering with 200+ brands, expected to reach 10M+ devices by 2026. Building the world's largest embodied intelligence data collection center, targeting 5M hours of high-quality data within one year.
Source: Toutiao
Editor's Take: Skip the hardware, build the data moat first — JD's "attached intelligence" approach is a uniquely pragmatic commercialization path. The data barrier from 10M+ device integration should not be underestimated.
11. ChinaJoy 2026: Unitree G1 Boxing, LimX Luna Catwalk, Dozens of Robot Companies Showcase
Date: 2026-07-31
Summary: ChinaJoy 2026 opens under the theme "Gaming with AI" — ~900 exhibitors, 140,000+ sqm. Unitree G1 boxing robots demo live combat, LimX Dynamics' 160cm entertainment robot Luna performs catwalk and gymnastics, Qiyuan debuts "world's first DIY-character personal robot" Q1 Explorer. Embodied intelligence makes its debut as "designer toy" to young consumers.
Source: Multiple tech media
Editor's Take: From factories to exhibitions to gaming conventions — humanoid robots are transitioning from "B2B tools" to "consumer products," an important signal.
12. Magic Atom HQ Factory Breaks Ground in Wuxi, 10,000-Unit Annual Capacity
Date: 2026-08-01
Summary: Magic Atom HQ factory officially breaks ground in Wuxi Liangxi Tech City. 4 joint module production lines + 2 full-assembly lines planned. At capacity: 9,000 small quadrupeds + 1,000 large quadrupeds annually. At WAIC 2026, globally debuted flagship MagicBot X1, with self-developed Magic-VLA K02 achieving >90% success rate on long-horizon manipulation tasks.
Source: Wuxi Bobao / Toutiao
Editor's Take: From "lab prototype" to "10,000 units/year" — Magic Atom's production line marks humanoid robots entering the industrialization phase.
13. Unitree Technology to Launch STAR Market IPO Subscription on August 5
Date: 2026-08-01
Summary: Quadruped robot leader Unitree Technology begins preliminary inquiry on August 5, with official online subscription on August 10 for STAR Market listing — among the first publicly listed humanoid robot companies.
Source: Multiple financial media
Editor's Take: Unitree's STAR Market listing will provide the first public-market pricing anchor for the Physical AI sector — one of August's most critical events.
14. Nearly 20 Automakers Collectively Enter Humanoid Robot Manufacturing
Date: 2026-08-01
Summary: Nearly 20 global automakers (including BMW, Mercedes-Benz, Hyundai, BYD, XPeng, Li Auto) have entered humanoid robots. XPeng IRON has begun small-batch trial production in Guangzhou factory. Smart vehicles and humanoid robots share >70% technology overlap and highly shared supply chains.
Source: Toutiao
Editor's Take: Automakers making robots isn't diversification — it's a natural extension of their tech stack. Global humanoid robot shipments expected to exceed 50,000 units in 2026.
15. Zhiyuan Yuanzheng A3 Ultra Selected as WAIC 2026 "Treasure of the Exhibition"
Date: 2026-07-31
Summary: Zhiyuan Robotics' commercial full-size Yuanzheng A3 Ultra (174cm, 700 TOPS compute) selected as one of WAIC 2026's "Top 10 Treasures" — the world's first mass-producible commercial full-size humanoid robot. On the same day, ZTE unveiled its Star-series humanoid robot and embodied AI solution.
Source: Jiemian / Cailian Press
Editor's Take: "Mass-producible commercial" — the most important qualifier in the humanoid robot sector. Zhiyuan and ZTE debuting on the same day signals accelerating competition.
IV. Autonomous Driving
16. Baidu Apollo Go Obtains Hong Kong and London Autonomous Driving Licenses
Date: 2026-07-31
Summary: Baidu's Apollo Go (LuoBoKuaiPao) obtains autonomous driving operation licenses in Hong Kong and London. The global autonomous driving competition officially expands beyond left-hand-drive markets (China and US) into right-hand-drive markets covering many developed economies. Initial deployment: 100 vehicles in designated areas.
Source: Multiple sources
Editor's Take: Hong Kong and London — two "tickets" with profound significance. This is China's autonomous driving technology breaking into the regulatory frameworks of developed economies.
17. Cao Cao Mobility Enters Autonomous Driving, Mass-Production Eva Cab by 2027
Date: 2026-07-31
Summary: Cao Cao Mobility enters the autonomous driving space, with mass-production model Eva Cab expected by 2027. Currently in early stages, but signals ride-hailing platforms' transformation toward autonomous mobility operators.
Source: Multiple sources
Editor's Take: A ride-hailing platform building its own autonomous vehicles — Cao Cao's path creates a differentiated competitive landscape against Apollo Go.
18. China's Embodied Intelligence H1 Financing Breaks RMB 93.5B, Up ~5x YoY
Date: 2026-07-31
Summary: H1 2026 embodied intelligence financing in China reaches RMB 93.5 billion (~$13B), up ~5x year-over-year, across 322 investment events. WAIC 2026 showcases 208 embodied AI terminals and 300+ real robots.
Source: IT Juzi / Investment China / East Money
Editor's Take: RMB 93.5B in six months — capital is voting with its feet. Embodied intelligence has become AI's largest capital reservoir.
19. Hong Kong Stock Market 2026 Equity Financing Reaches $67B, AI-Led
Date: 2026-07-31
Summary: Hong Kong's 2026 cumulative equity financing reaches $67 billion, exceeding 70% of the entire 2025 total. IPOs account for $30.3B, with AI-related sectors as the primary driver.
Source: Cailian Press
Editor's Take: $67B in financing — the Hong Kong stock exchange is becoming the preferred listing venue for AI companies.
V. World Model / Physical AI
20. August: "Robot Super Month" — Unitree IPO + Shanghai Embodied AI Expo + Beijing World Robot Conference
Date: 2026-08-01
Summary: August is dubbed "Robot Super Month": Unitree STAR Market subscription (Aug 5), Shanghai Embodied Intelligence Robot Industry Conference (Aug 12-14), 2026 World Robot Conference (Aug 19-23, Beijing: 300+ exhibitors, 2,000+ exhibits, 150+ global debuts). Multiple top brokerages publish bullish research reports on Physical AI.
Source: Toutiao / Multiple brokerages
Editor's Take: IPO + two major expos in one month — August will be the "pricing month" for Physical AI.
21. World Robot Conference to Establish SOE Robot Innovation Consortium
Date: 2026-08-01
Summary: The 2026 World Robot Conference (Aug 19-23, Beijing) will announce the establishment of a State-Owned Enterprise Robot Innovation Consortium, uniting ~50 central SOEs to bridge scenario demand with technology R&D, potentially releasing massive industrial robot procurement demand.
Source: Toutiao
Editor's Take: 50 SOEs united — potentially the single largest B2B robot demand signal in China.
22. Physical AI Industry Cycle Shift: From Digital AI to Physical AI
Date: 2026-08-01
Summary: Multiple brokerage research reports note the AI industry is shifting from "Digital AI" (LLMs/multimodal generation) to "Physical AI" (embodied intelligence/humanoid robots/autonomous driving). Four foundational technologies converging: world models, VLA end-to-end models, declining hardware costs, and structural labor shortages. 2026-2028 is the critical transition period.
Source: Toutiao / Multiple brokerage reports
Editor's Take: Digital AI's marginal elasticity is narrowing — Physical AI is the highest-growth tech sector thesis right now.
23. Shenji Completes ~RMB 3B Financing Round
Date: 2026-07-31
Summary: Autonomous driving chip company Shenji completes ~RMB 3 billion in new financing, injecting momentum into the autonomous driving and robotics chip sector.
Source: CSDN / IT Juzi
Editor's Take: RMB 3B financing — domestic smart driving chip sector continues to heat up.
24. Kuaishou Kling AI Raises $3B, Largest Single Video Model Funding Globally
Date: 2026-07-31
Summary: Kuaishou's Kling AI raises $3 billion — the largest single financing event in the global video generation model sector. Kling AI continues to lead in video generation.
Source: CSDN / Investment China
Editor's Take: $3B — video model capital enthusiasm remain strong, but commercialization validation is the next critical test.
25. DeepSeek Valued at RMB 480B, Preparing for A-Share IPO
Date: 2026-07-31
Summary: DeepSeek's valuation reaches RMB 480 billion (~$66B), with A-share IPO preparations underway. As China's benchmark AI model company, its listing will be a landmark event in China's AI capital markets.
Source: CSDN / Investment China
Editor's Take: RMB 480B valuation + A-share IPO — DeepSeek's listing will be the defining moment for China's AI industrial capitalization.
26. Wuxi Targets RMB 20B Embodied Intelligence Industry Revenue
Date: 2026-08-01
Summary: Wuxi aims to become a national benchmark for embodied intelligence, targeting RMB 20B+ industry revenue and 5,000+ humanoid robot mass production in 2026. 13 companies including Galaxy General, Zhiyuan, and Standard Robots have established operations; 178 supply chain companies; Q1 revenue up 23.5% YoY.
Source: Wuxi Bobao
Editor's Take: 178 supply chain companies, 5,000-unit production target — Wuxi is becoming China's "Silicon Valley" for Physical AI.
27. Perovskite Solar Cell Efficiency Reaches 25.53%
Date: 2026-08-02
Summary: Perovskite photovoltaic lab conversion efficiency reaches 25.53%, with 95% performance retention after aging — further clearing long-term durability barriers to commercialization.
Source: Toutiao
Editor's Take: 25.53% + 95% retention — perovskite takes another step toward mass production, positive for energy AI supply chain.
28. TCL CSOT Builds Printed OLED Production Line
Date: 2026-08-02
Summary: TCL CSOT builds a printed OLED production line with significantly improved material utilization, reducing large-screen manufacturing costs. Indirect benefit to AI display terminal supply chain (robot vision, AR/VR).
Source: Toutiao
Editor's Take: Printed OLED cost reduction positively impacts screen costs for AI terminal devices.
Sector Statistics
| Sector | Count |
|---|---|
| AI Foundation Models | 7 |
| AI Software | 6 |
| Humanoid Robots | 6 |
| Autonomous Driving | 4 |
| World Model / Physical AI | 9 |
| Total | 28 |
🔔 East Voice — Physix Frontier's China AI column for global readers
East Voice Westward, Chinese Perspective.
© 2026 Physix Frontier · East Voice