社区讨论 · 政策

East Voice · China AI Daily | August 3, 2026 · Issue #762

EastVoice8月3日2026/08/03 274 浏览

August 3, 2026 · Monday · Issue #762

Editor’s Note

Chinese AI models sweep the top 5 of global API call rankings — an unprecedented milestone in the open-source ecosystem’s journey from catching up to leading.

Three headlines dominate this edition: First, the OpenRouter leaderboard is entirely Chinese — Xiaomi’s MiMo-V2.5 leads, followed by DeepSeek and Tencent Hunyuan entries, while Alibaba’s Qwen series surpasses 1 billion downloads on Hugging Face. Second, Moonshot AI closes a $3.5 billion oversubscribed Series F (3x oversubscribed) as Kimi K3 becomes the largest open-weight model ever at 2.8 trillion parameters — demand is so overwhelming that new user registrations are paused due to GPU constraints. Third, ByteDance executes its biggest ToB restructuring since 2021, merging Feishu, Doubao, and Volcengine into a unified platform.

Editor’s Note: Twelve months ago, US models commanded 72% of OpenRouter’s call volume. Today, Chinese models own the top five spots. But “open-source leadership” is not the same as “commercial viability” — converting ecosystem dominance into sustainable revenue is the next defining challenge. — East Voice Editorial

Track 1: LLM / Foundation Model

1. Chinese AI Models Dominate Global API Call Rankings: Top 5 on OpenRouter All Chinese-Made

Date: 2026-08-03
Summary: Xiaomi’s MiMo-V2.5 tops OpenRouter’s latest global model call volume rankings with 10.5 trillion tokens processed weekly, followed by two DeepSeek models and Tencent Hunyuan Hy3 — all five spots held by Chinese developers.
Source: Toutiao (Jiuyi)
Editor’s Take: From 72% US dominance to a complete Chinese sweep in just 12 months — this is not merely a market share shift but a fundamental redistribution of global AI industry power.

2. DeepSeek V4-Flash Official API Launches with 6x Agent Performance Surge

Date: 2026-08-01
Summary: DeepSeek launches V4-Flash official API — same 284B-parameter MoE architecture, but post-training re-tuning propels DeepSWE coding agent score from 7.3 to 54.4, with cache-hit pricing at ¥0.2 per million tokens (1/90th of Claude’s equivalent).
Source: Toutiao / Malpass / Top AI Stories
Editor’s Take: One developer ran 323 million tokens for $4.55. The marginal returns of “same architecture, better post-training” have been massively underestimated — autonomous coding agents just got dramatically cheaper.

3. Tencent Hunyuan Hy3 Goes Open-Source Under Apache 2.0 with No Geographic Restrictions

Date: 2026-08-01
Summary: Tencent releases Hunyuan Hy3: 295B total params / 21B active, MoE with fast/slow thinking, 256K context window. API calls surged 68x vs Hy2, topping OpenRouter’s global leaderboard. Apache 2.0 license — notably more permissive than DeepSeek’s MIT or Moonshot’s modified MIT.
Source: BigHat Group / Toutiao
Editor’s Take: Tencent merges its LLM and multimodal divisions into a unified Foundation Model Department — a clear signal that the agentic era demands tight integration across perception, reasoning, and execution.

4. Moonshot AI Closes $3.5 Billion Series F at $35 Billion Valuation — 3x Oversubscribed

Date: 2026-08-02
Summary: Moonshot AI completes a $3.5B+ Series F round, valued at $35B, oversubscribed 3x and closed early — the largest single AI funding round in China for 2026.
Source: Titanium Media
Editor’s Take: 3x oversubscription = capital voting with conviction. But the GPU bottleneck (paused registrations) proves that model capability and user demand are solved — the binding constraint is purely on the supply side.

5. MiniMax H3 Tops Global Video Editing Benchmarks, Open-Source Release Imminent

Date: 2026-08-02
Summary: MiniMax H3 scores 1130 Elo to top Artificial Analysis’s video editing leaderboard, supports native 15-second 2K with stereo audio at ¥0.8/second (1/3 the cost of competitors). Model weights to be open-sourced within days. Stock surged 13% on the day.
Source: Yilan Business / AGI HUNT
Editor’s Take: First time a Chinese model tops a global video editing benchmark. The “top the charts, then open-source” strategy is precise — use benchmark credibility to build trust, use open-source to scale the ecosystem.

Track 2: AI Software / SaaS

1. ByteDance Merges Feishu + Doubao + Volcengine: Biggest ToB Restructuring Since 2021

Date: 2026-08-02
Summary: ByteDance issues internal memo merging Feishu product team with Doubao product team, and Feishu GTM with Volcengine into a “Creativity Services Platform,” unified under Tan Dai. ARR has reached $4 billion.
Source: Titanium Media
Editor’s Take: This is ByteDance’s largest ToB restructuring since 2021. Unifying MaaS (Model-as-a-Service) and SaaS means ByteDance no longer sells “tools” and “models” separately — integration is the sprint.

2. Tesla China Infotainment Integrates Doubao LLM; Qwen Simultaneously in Beta

Date: 2026-08-01
Summary: Tesla China’s v2026.14.13 OTA pushes Doubao large model to Model 3/Y/S/X for voice and contextual understanding; Alibaba’s Qwen completes cabin beta testing in select new vehicles.
Source: Jiemian News / Toutiao
Editor’s Take: Chinese LLM providers are bypassing the “cloud chatbot” paradigm to embed low-latency hybrid inference directly into automotive supply chains. Whoever wins Tesla China’s default slot captures a million-level DAU entry point.

3. Tencent ADP 4.0 Global Launch + WorkBuddy App on iOS/Android/HarmonyOS

Date: 2026-08-01
Summary: At WAIC, Tencent unveils Agent Development Platform 4.0 with LINE/Telegram/Google Workspace integrations; standalone WorkBuddy app launches on three platforms; QQ Pet returns with Hy3-powered 3D interactive pets.
Source: BigHat Group
Editor’s Take: ADP 4.0 targets overseas (LINE + Telegram + Google Workspace trifecta), WorkBuddy targets domestic — Tencent’s agent strategy runs dual tracks. The QQ Pet revival shows “emotional AI + nostalgia IP” is a low-cost C-end agent validation path.

4. ByteDance Seedance 2.5: 30-Second Video Generation with 50 Multimodal Reference Inputs

Date: 2026-08-02
Summary: ByteDance Dreamina launches Seedance 2.5 globally — supports up to 50 multimodal reference inputs and 30-second single-generation video at industry-lowest access costs.
Source: AGI HUNT
Editor’s Take: 50 reference inputs = precise style and detail control; 30 seconds covers most short-video scenarios. ByteDance’s “low price + high quality” combination continues pressuring competitors in video generation.

5. Baidu Miaoda 3.5: AI No-Code Platform Reaches 35 Million Users

Date: 2026-08-03
Summary: Baidu’s Miaoda has served 35 million users and created 3.5 million applications. Version 3.5 adds SEO Agent, iOS packaging, shared backend, and five other capabilities. H1 2026 market share: 33.4% (#1 in China’s AI-native no-code market by MAU).
Source: 36Kr
Editor’s Take: “Built an AI comic app in a week, earned ¥120K” / “Oil field engineer’s system saved ¥1.4M” — Miaoda proves AI no-code is not a toy but a real productivity tool. But the “production-grade interaction platform” positioning depends on ecosystem vitality.

6. Multi-Model Tiered Routing Becomes New Standard for AI Applications

Date: 2026-08-02
Summary: Orient Securities research note: the era of one premium flagship model covering all tasks is ending. Multi-model on-demand routing is becoming mainstream — simple tasks use lightweight models, complex tasks use flagships, reducing overall cost by 50-90%.
Source: STAR Market Daily
Editor’s Take: “Not every question needs the smartest brain” — tiered routing makes AI application economics finally work. For model vendors: the lightweight model market may be larger than the flagship market.

7. Model Price Cuts + Open-Source Drive AI Democratization, Commercialization Accelerates

Date: 2026-08-02
Summary: Multiple brokerage research notes conclude that DeepSeek V4-Flash, OpenAI Luna’s 80% price cut, and Kimi K3’s open-source release collectively mark the global LLM transition from “closed-source premium” to “utility-grade infrastructure.”
Source: STAR Market Daily
Editor’s Take: Cheaper and better models — the real commercial value is in the application layer. ByteDance Doubao, Tencent WorkBuddy, and Alibaba Qwen are rapidly accumulating users and habits. “Whoever makes users dependent first” wins the application race.

8. Alibaba Bans Claude Code Company-Wide; Chinese Government Warns of “Spyware” Risk

Date: 2026-08-01
Summary: Alibaba imposes a company-wide ban on Anthropic’s Claude Code citing “spyware concerns.” China’s government separately warns that Claude Code poses “a serious threat” to Chinese users.
Source: BigHat Group
Editor’s Take: The AI cold war extends from the model layer to developer tooling. Anthropic accuses Chinese firms of model distillation → Alibaba bans Claude Code → Chinese government intervenes. This is reshaping global developer toolchain choices.

Track 3: Humanoid Robot

1. Magic Atom Breaks Ground on Wuxi HQ, Plans 10,000+ Units/Year Production

Date: 2026-08-01
Summary: Magic Atom’s national headquarters breaks ground in Wuxi Liangxi Tech City with 4 joint module lines + 2 assembly lines. Target capacity: 9,000 small quadrupeds + 1,000 large quadrupeds annually. MagicBot already deployed at Dreame factories.
Source: Toutiao
Editor’s Take: The “10,000-unit cost curve” for humanoid robots is, for the first time in China, written into a construction schedule. Physical AI platforms are transitioning from demos to production.

2. China Robotics Competition: Full-Size Humanoids Complete Industrial Tasks Without Remote Control for First Time

Date: 2026-08-01
Summary: The 28th National Robotics Competition’s humanoid track in Beijing sees 213 teams with 350+ robots. Full-size category goes fully un-teleoperated for the first time, replicating high-risk inspection, production line loading/unloading, and auto parts sorting in a 1:1 industrial setting.
Source: Toutiao / CCTV
Editor’s Take: The benchmark shifts from “impressive movements” to “achievable industrial KPIs” — 50 minutes of autonomous slope climbing, obstacle avoidance, and sorting. This is a generalization stress test for manufacturers.

3. SenseTime Open-Sources ACE-Brain Embodied Foundation Model for All Robot Categories

Date: 2026-08-01
Summary: SenseTime officially open-sources ACE-Brain, an embodied foundation model compatible with all robot categories, significantly lowering the barrier for domestic embodied intelligence development.
Source: Toutiao / Robot Zero
Editor’s Take: “One brain for all bodies” — open-sourcing an embodied foundation model is analogous to Android for phones: hardware vendors focus on physical differentiation while the software layer is shared for cost reduction.

4. Agibot Yuanzheng A3 Achieves 10,000-Unit Delivery, Enters Auto Parts Factories at Scale

Date: 2026-08-01
Summary: Agibot’s Yuanzheng A3 series achieves sustained 10,000-unit-level delivery, deployed at scale in automotive parts factories for material transport and flexible assembly. Proprietary WITA embodied model continues iterating.
Source: Toutiao / Robot Zero
Editor’s Take: “10,000-unit delivery” is no longer a slogan — Agibot is among the first in China to achieve normalized industrial humanoid operations. Production data becomes the core asset for future funding and customer trust.

5. Unitree Robotics Enters Critical Phase of STAR Market IPO

Date: 2026-08-01
Summary: Unitree Robotics’ STAR Market IPO enters the critical pre-subscription phase. G1 humanoid robots continue deploying in logistics and power grid inspection while providing hardware platforms for NVIDIA GR00T.
Source: Toutiao / Robot Zero
Editor’s Take: Unitree’s unique positioning: “best body, no brain” — they make the hardware, let others provide the intelligence. Supplying NVIDIA GR00T means global robotics developers can pair Unitree bodies with any AI brain.

6. BYD Confirms August Humanoid Robot Debut, Leveraging Automotive Supply Chain

Date: 2026-08-01
Summary: BYD confirms a new humanoid robot will officially debut in August, initially for showroom service and interactive demonstrations, with automotive supply chain advantages fully applied to robotics.
Source: Toutiao / Robot Zero
Editor’s Take: What does BYD’s entry mean? Automotive-grade supply chain management + mass manufacturing experience + proprietary showroom scenarios. If automotive cost discipline transfers to robotics, it could be a category disruption.

7. UBTECH U1 Pre-Orders Exceed 13,000 Units

Date: 2026-08-01
Summary: UBTECH’s U1 series surpasses 13,000 pre-sale orders. Walker humanoid continues进驻 automotive production lines for flexible operation testing. Commercialization accelerates.
Source: Toutiao / Robot Zero
Editor’s Take: 13,000 pre-orders is a significant milestone for consumer humanoid robots — purchase intent exceeds industry expectations, but delivery experience and long-term engagement are the real tests.

8. Leisai Intelligence Raises ¥1.144B for Expansion; Frameless Torque Motor Orders Exceed 1 Million

Date: 2026-08-01
Summary: Leisai Intelligence’s ¥1.144 billion secondary offering approved to build 7.5 million annual capacity for motion control components. Frameless torque motor backlog exceeds 1 million units in H1, covering 80% of major domestic humanoid manufacturers.
Source: Robot Industry Chain Daily / Toutiao
Editor’s Take: Core components are profitable before complete machines — 80% market share makes Leisai the “hidden champion” of China’s robot supply chain. Their capacity expansion directly dictates OEM delivery timelines.

Track 4: Autonomous Driving

1. Baidu Apollo Go RT6 Launches London Road Trials — Robotaxi Expands into Europe

Date: 2026-08-02
Summary: Baidu and Lyft subsidiary Freenow begin public road testing of 6th-gen Apollo Go RT6 in London’s Brent borough. Right-hand-drive vehicles with safety operators; commercial rollout planned for 2027, pending TfL and CCAV approval.
Source: AUTO Connected Car
Editor’s Take: From Hong Kong Airport Island to central London — Apollo Go’s overseas path upgrades from “Asian pilot” to “European deployment.” Baidu’s L4 perception stack + Freenow’s local dispatch is a cross-border division of labor worth watching.

2. XPeng IRON Humanoid Robot Enters Small-Batch Trial Production in Guangzhou

Date: 2026-08-01
Summary: XPeng’s IRON humanoid robot enters small-batch trial production at its Guangzhou factory, with mass production progress exceeding expectations. Autonomous driving technology extends into robotics.
Source: Toutiao / Robot Zero
Editor’s Take: XPeng repurposes autonomous driving perception and decision-making for humanoid robots — “the tech that drives cars can build people.” If mass production beats expectations, XPeng could be the first automaker to commercially deploy humanoid robots.

3. Tesla FSD v14: Speed Profiles and Arrival Options Go Live

Date: 2026-08-01
Summary: Tesla FSD (Supervised) v14 adds Speed Profiles (driver-configurable speed offsets) and Arrival Options (endpoint parking preferences). Urban unprotected left-turn success rates significantly improved.
Source: Tesla Official / CSDN
Editor’s Take: FSD evolves from “can it drive” to “does it drive the way you like.” Personalized driving configuration signals the system is mature enough for fine-grained experience tuning. The $99/month subscription model is becoming mainstream.

4. China H1 Humanoid Robot Production Exceeds 40,000 Units; Full-Year Target: 100,000+

Date: 2026-08-01
Summary: MIIT and SASAC joint initiative sets a year-end target of 10,000+ robots in normalized deployment. Latest data shows H1 domestic humanoid robot production exceeded 40,000 units, with full-year output expected to surpass 100,000.
Source: Toutiao / Robot Zero
Editor’s Take: 40,000 in H1 → 100,000 for the year — these numbers mean humanoid robots have transitioned from “lab artifacts” to “industrial products.” The late-August World Robot Conference in Beijing will be the showcase window.

5. MIIT Sets 2026 Year-End Target: 10,000+ Humanoid Robots in Routine Deployment

Date: 2026-08-01
Summary: China’s Ministry of Industry and SASAC jointly mandate: 10,000+ humanoid robots in normalized deployment by end of 2026. Late August World Robot Conference in Beijing to feature live hardware demonstrations.
Source: Toutiao / Robot Zero
Editor’s Take: Official designation of “Year of Scale” — when policy targets align with production data (40K in H1), the industrialization of humanoid robots enters an acceleration phase.

Track 5: World Model / Physical AI

1. Tsinghua Team Publishes Centimeter-Scale Morphing Robot in Science Advances

Date: 2026-08-01
Summary: Tsinghua’s Zhang Yihui team achieves centimeter-scale morphological and dynamic control via musculoskeletal actuators, creating a 1.8g quadruped-to-humanoid-switching robot and a 6.4g gliding squirrel-like multimodal robot.
Source: DeepTech / Toutiao
Editor’s Take: Sub-10-gram, shape-shifting, gliding — pushing the “smallest autonomous unit” to insect scale. Pipeline inspection and post-disaster micro-space exploration could commercialize before large-format humanoids.

2. Qiyuan Robot Debuts Q1 Explorer / T1 at ChinaJoy with World of Warcraft Skins

Date: 2026-08-01
Summary: Qiyuan Robot (subsidiary of Shangwei Xincai) launches at ChinaJoy with DIY-customizable Q1 Explorer and transformable T1, including limited-edition World of Warcraft character skins.
Source: Toutiao
Editor’s Take: Consumer embodied intelligence’s “skin-first” strategy — earn emotional premiums through IP co-creation. But it also exposes the core tension: interaction depth lags behind cosmetic iteration speed.

3. JD.com Goes All-In on Embodied Intelligence, AI Models to Power Millions of Existing Devices

Date: 2026-08-01
Summary: JD.com officially announces full commitment to embodied intelligence — not competing on hardware but powering millions of existing smart devices with AI models, building a large-scale physical intelligence ecosystem.
Source: Toutiao / Robot Zero
Editor’s Take: “Don’t build bodies — build souls.” JD.com’s play is installing AI brains into existing devices. The warehouse logistics scenario provides a natural training ground and monetization pool with millions of deployed units.

4. Snapdragon 2nd-Gen Smart Ring Debuts at ChinaJoy: Cross-Device Gesture + AR + PC Loop

Date: 2026-08-01
Summary: Qualcomm’s Snapdragon pavilion at ChinaJoy showcases the 2nd-gen smart ring paired with AR glasses and Snapdragon PC — positioned as an “always-on, low-power intent capture device.”
Source: Toutiao
Editor’s Take: Ring = micro-interaction input beyond voice. Wearable AI moves from “earbuds dominate” to multimodal sensor fusion — when a ring captures gestures and intent, “natural interaction” is redefined.

5. NVIDIA Vera CPU Enters Mass Production: From “Selling GPUs” to “Selling AI Factories”

Date: 2026-08-01
Summary: NVIDIA Vera CPU officially enters mass production and delivery, forming a full-stack vertical solution with GPU, BlueField networking, and NVLink switches. Cloud providers can directly access “plug-and-compute” AI factories.
Source: Jixin / Toutiao
Editor’s Take: NVIDIA is transforming from “chip vendor” to “turnkey solution provider” — OEM bargaining power shifts toward NVIDIA, compressing whitebox server and x86 vendors’ market space. Chinese smart computing center tenders now implicitly require “Vera-class full-stack delivery.”

6. Insta360 GO Ultra Previews AI Assistant “Kira” — Thumb Camera Evolves into Personal AI Companion

Date: 2026-08-01
Summary: Insta360 founder Liu Jingkang teases GO Ultra’s “Hey, Kira” (launch set for August 5), claiming potential to become “the largest personal AI assistant outside Meta by installed base and shipments.”
Source: Toutiao
Editor’s Take: Thumb camera + AI assistant = first-person continuous perception + local summarization + proactive alerts. If delivered, the AI hardware battlefield expands from glasses/earbuds to “clip-on micro-camera-voice terminals.”

7. Research: 78% of Agent Task Failures Are “Silent Errors”

Date: 2026-08-02
Summary: Researchers discover 78% of agent task failures are “silent errors” — tool execution succeeds but domain policies are violated, undetectable by the agent itself. Paper accepted at KDD 2026 Workshop.
Source: arXiv / Toutiao
Editor’s Take: “The tool reported success, but business logic was violated” — this is the biggest reliability threat to agent systems. KDD 2026 acceptance signals academic recognition: agent error detection cannot rely solely on tool return values.

8. Sixunion AI PC Full Lineup at ChinaJoy: 128GB Unified Memory as Edge Agent Watershed

Date: 2026-08-01
Summary: At ChinaJoy, Sixunion showcases a full AI PC matrix with AMD Ryzen AI processors — high-end workstations support 128GB unified memory and multi-machine clusters, enabling parallel multi-agent operation locally.
Source: Toutiao
Editor’s Take: 128GB unified memory = local 7B-13B parameter models + parallel multi-agent execution. “Agents that work offline” become a new government and enterprise procurement standard — driven by dual demands for data security and computational sovereignty.

Track Statistics

Track	Count
LLM / Foundation Model	6
AI Software / SaaS	7
Humanoid Robot	8
Autonomous Driving	5
World Model / Physical AI	8
Total	34

东声 East Voice — Physix Frontier’s China AI column for global readers

东声西达，中国视角。

政策