China AI Daily

July 14, 2026 (Tuesday) ｜ Team B · China → Global

Physix Frontier · bbs.physixfrontier.com

---

Editor’s Note

Today’s theme: China’s AI is shifting from “model chasing” to full-stack autonomy. Tencent’s HunYuan Hy3 crashes its own servers on launch day — a 295B-parameter MoE model with 21B active parameters, free for all users via WorkBuddy. MiniMax closes a massive HK$16B funding round (7x oversubscribed) while secretly developing a 2.7-trillion-parameter M3 Pro. DeepSeek V4 drops July 15 with China’s first-ever peak-valley API pricing. Unitree Robotics gets STAR Market IPO approval at a 42B RMB valuation — the only profitable humanoid robot company going public. AgiBot rolls its 15,000th unit off the line. MIIT forecasts full-year humanoid robot production will break 100,000 units. XPeng runs the full Robotaxi chain in 8 months with 3,000 TOPS (world’s highest on-vehicle compute). And Momenta lists on HKEX with 414x oversubscription — then closes flat on day one. The message is clear: China’s AI stack is going vertical, from chips to models to robots to roads.

---

Track 1: AI Models (LLM / Foundation Models / Reasoning)

1 | Tencent HunYuan Hy3 Launches and Immediately Overwhelms Its Own Servers

Tencent officially released HunYuan Hy3 (混元3), a 295B-parameter MoE model with 21B active parameters and 256K context window. On July 8, demand on Tencent’s WorkBuddy platform spiked so hard that compute resources were maxed out — afternoon queue rates exceeded 50%. Tencent scrambled to add capacity overnight. Hy3 is now integrated across WorkBuddy/CodeBuddy, Yuanbao, Marvis, and ima. This is the first major release under new AI lead Shunyu Yao (appointed December 2025, reporting directly to President Martin Lau).

Sources: Tencent Official, AI Weekly Observer

---

2 | MiniMax Closes HK$16B Funding Round, 7x Oversubscribed — CEO Vows No Salary Until AGI

MiniMax completed a HK$16 billion (~$2B) equity financing comprising HK$9.54B in share placement (35.6M shares at HK$267.99 each) plus HK$6B in convertible bonds. Over 100 institutions participated with 7x coverage, including 20+ sovereign wealth and long-only funds. CEO Junjie Yan announced he will take zero salary until the company achieves AGI, and committed 4% of his personal equity to team incentives plus 1% to open-source community support. Over 80% of pre-IPO and cornerstone investors have pledged to hold long-term.

Sources: East Money, IT Home

---

3 | MiniMax Developing 2.7-Trillion-Parameter “M3 Pro” — Targeting Q3 Release

Reports confirm MiniMax is building M3 Pro, a next-generation LLM with 2.7 trillion parameters — designed to close gaps in long-horizon task processing and complex reasoning. Expected to launch and open-source as early as Q3 2026. This positions MiniMax among the top 3 Chinese foundation model companies by parameter scale.

Sources: China Securities Journal, IT Home

---

4 | DeepSeek V4 Confirmed for July 15 Launch — China’s First Peak-Valley API Pricing

DeepSeek V4 officially launches July 15 with two landmark changes: (1) China’s first peak-valley pricing — off-peak API at just ¥0.5/million tokens, peak hours (9:00-12:00, 14:00-18:00) at 2x; (2) context window expanded from 128K to 1M tokens, with 40-60% throughput improvement. DeepSeek V4 Pro API has already been permanently reduced to 1/4 of original price. Meanwhile, Moonshot AI’s Kimi K3 (reportedly 2.5T parameters) is expected later this month, directly targeting V4 Pro.

Sources: AI Daily, DeepSeek Official

---

5 | Meituan Open-Sources LongCat-2.0: First Trillion-Parameter Model Trained Entirely on Domestic Chips

Meituan officially open-sourced LongCat-2.0, a 1.6 trillion parameter model (~48B active) — the first trillion-parameter model to complete its entire training and inference lifecycle on a 50,000-card domestic computing cluster. Key innovations include LongCat Sparse Attention (linear complexity for million-token contexts), N-gram Embedding (135B parameters for token representation), and native 1M context window. Benchmark highlights: SWE-bench Pro 59.5 (surpassing Gemini 3.1 Pro 54.2, GPT-5.5 58.6). Available on HuggingFace, GitHub, and ModelScope with optimized inference for both GPU and Huawei Ascend NPU.

Sources: Meituan Tech Blog, 36Kr, AIToolly

---

6 | Chinese AI Models 60-90% Cheaper Than US Counterparts — Rapidly Gaining US Enterprise Adoption

Chinese frontier models are now 60-90% cheaper than OpenAI/Anthropic equivalents, with a capability gap of only 6-9 months. On OpenRouter, US enterprise usage of Chinese models has consistently exceeded 30% of tokens since February 2026, peaking at 46% (vs. 11% average in the prior 12 months). Zhipu’s GLM-5.2 saw daily token calls surge 27x and customer count grow 80x in its first week. US startup Lindy reported saving millions after switching all AI operations to DeepSeek.

Sources: CCTV Finance, OpenRouter

---

7 | China’s AI Industry to Grow 30%+ in 2026; AI Phone/PC Sales to Surpass Non-AI Products

NDRC disclosed at the WAIC press conference that China’s AI core industry exceeded ¥1 trillion in 2025, with projected growth of 30%+ in 2026. Over 30 national AI application pilot bases have been established, with SOEs opening 1,000+ application scenarios. AI penetration in key industries has surpassed 80%. AI smartphone and AI PC shipments exceeded 100M units in 2025, with 2026 expected to see AI products outsell non-AI products for the first time.

Sources: NDRC, WAIC Press Conference

---

Track 2: AI Software (Applications / Tools / Platforms)

1 | Alibaba Bans Claude Code Company-Wide — 120,000 Developers Migrate to Domestic Tools

Alibaba’s internal directive took effect July 10: complete ban on all Anthropic products (Sonnet, Opus, Fable, Claude Code) across all office networks and devices. The trigger: developer community reverse-engineered Claude Code v2.1.91+ and found a hidden user detection mechanism using Unicode apostrophe variants to distinguish Chinese domains and AI lab accounts. MIIT had previously flagged Claude Code as having “serious security backdoor risks.” Alibaba recommends migration to its in-house Qoder platform (5M+ global users). Tongyi Lingma launched a “one-click migration” tool, and Qwen-Coder tripled its free quota. An estimated 120,000 developers face mandatory migration.

Sources: AI Daily, MIIT, Alibaba Internal Memo

---

2 | WeChat Rolls Out Major AI Feature Updates

WeChat published 8 major June updates: developer platform now offers auto and developer modes for AI ecosystem integration; WeChat Pay AI toolkit upgraded to v2.0 with 9-language native support and Mermaid-format docs (50% token reduction); QQ Mail begins “Agent Mail” beta supporting 5 mainstream agents; WeCom 5.0.9 launches “Service Summary” with native AI assistant “Dayuan” in gray beta; WeChat Input Mac upgrades voice input model; Official Account comments now support 9-image posts; WeChat Read adds one-click reading summary skill.

Sources: WeChat Open Platform

---

3 | Alipay AI Open Platform Goes Live — Inviting Enterprise Developers

Ant Group launched the Alipay AI Open Platform on July 7, targeting enterprise developers and ISVs. Built on a three-layer architecture (model capabilities → industry solutions → developer tools) with Ant’s in-house Bailing LLM, covering intelligent customer service, marketing copy generation, business analytics, and code assistance. Pricing: free base calls + pay-per-use for advanced capabilities. First-wave adopters include KFC, Mixue, Luckin Coffee, Gaode, and Didi.

Sources: Ant Group Official

---

4 | Moore Threads Launches MTT AICUBE: Home AI Hub at ¥9,999

Moore Threads began selling MTT AICUBE, its first consumer product — a “home AI hub” integrating AI agent + AI PC + AI NAS. Powered by its proprietary “Changjiang” SoC with 50 TOPS heterogeneous AI compute, 32GB unified memory, and 1TB all-flash SSD. Core agent “Xiaomai” supports voice cloning, avatar customization, and cross-app task execution with 95%+ success rate on high-frequency tool calls.

Sources: Moore Threads Official

---

5 | Chinese Banks Burn 140 Trillion Tokens Per Day — ROI Reckoning Begins

Multiple Chinese banks disclosed staggering token consumption: China Merchants Bank leads at 330 billion tokens/day (cost-to-revenue ratio 20%); Postal Savings Bank at 100B+ daily with 6M+ model calls across 370+ AI scenarios; WeBank at 50B+. National daily token consumption surged from 100B in early 2024 to 140 trillion by March 2026 — a 1,000x increase in two years. But banks are now pivoting to mandatory ROI assessment: CMB restricts coding AI to just 5% of compute (citing “spaghetti code” risks), and some banks are cutting minimum token KPIs by 20%.

Sources: Economic Daily, Securities Daily, EEO

---

6 | China’s First Systemic AI Governance Framework for Banking and Insurance

NFRA issued the “Guiding Opinions on Safe Development and Application of AI in Banking and Insurance” — China’s first comprehensive AI regulation for finance. Core principles: “whoever uses it is responsible,” autonomous and controllable platforms mandatory, no “AI for AI’s sake.” High-risk AI applications (credit, claims, trading) require risk management committee approval. Personal data prohibited from training generative AI. Algorithmic transparency and explainability standards required.

Sources: NFRA Official, Legal Daily

---

Track 3: Humanoid Robots (Embodied AI / Industrial Applications)

1 | Unitree Robotics Gets STAR Market IPO Approval — A-Share’s “First Humanoid Robot Stock”

Unitree Robotics (宇树科技) received STAR Market IPO registration after a record 104-day review — the fastest under the pre-review mechanism. IPO raises ¥4.2B (~$580M) at a valuation of ~¥42B (~$5.8B), with nearly ¥2B earmarked for robot AI model R&D. 2025 financials: revenue ¥1.708B (+335% YoY), non-recurring net profit ¥600M, gross margin 60.13% — the only profitable humanoid robot company at scale. 2025 shipments: 5,500 units (global #1). Morgan Stanley raised its 2026 China humanoid robot forecast from 28,000 to 50,000 units.

Sources: Securities Times, 36Kr

---

2 | AgiBot Hits 15,000 Units — From 6 Prototypes to Mass Production in Under 3 Years

AgiBot (智元机器人) announced its 15,000th general-purpose humanoid robot rolled off the production line. Growth trajectory: 6 prototypes (2023) → 10,000 (March 2026) → 15,000 (June 2026). The 5,000-unit increment took less than 3 months. 90% of components sourced from Yangtze River Delta supply chain. Robot models cover industrial handling, logistics sorting, and security patrols. Shanghai port exported ¥8.36B (~$1.2B) in robots in Jan-May 2026, shipping to 113 countries (40%+ of China’s total robot exports).

Sources: Xinhua, China Business News

---

3 | MIIT: China’s Humanoid Robot Production to Break 100,000 Units in 2026

At the WAIC press conference, MIIT disclosed that China’s AI core industry exceeded ¥1.2 trillion in 2025, and humanoid robot full-year production is expected to break 100,000 units in 2026. China has developed nearly 200 key AI standards, with model downloads ranking #1 globally. AtomGit and other open-source communities have aggregated over 11 million registered users.

Sources: MIIT, CCTV News

---

4 | Raythink (瑞为技术) Lists on HKEX as “First Visual Embodied AI Stock” — 3,646x Oversubscribed

Raythink officially listed on HKEX on July 8 as the “first visual embodied AI stock.” Hong Kong public offering was 3,646x oversubscribed. Revenue grew from ¥242M (2023) to ¥443M (2025), with smart aviation (¥172M), smart commercial (¥154M), and smart driving (¥116M) as key segments.

Sources: Science and Technology Weekly

---

5 | Ant LingBo Open-Sources Two Embodied AI Models — World Model Gets Agent Mechanism

Ant Group’s LingBo released LingBot-World 2.0 (first world model with Agent mechanism — supports 10-minute stable frames + WASD real-time interaction) and LingBot-Video (world’s first MoE-based open-source video generation foundation model for embodied AI, trained on 70,000 hours of embodied data). Both models lead open-source benchmarks in physical plausibility and action coherence.

Sources: AI Daily, Ant LingBo

---

6 | China’s Humanoid Robot Exports Surge 210% in Q1 2026

Q1 2026 data shows China’s humanoid robot exports grew 210% YoY. Jan-May 2026, robot industry enterprises above designated size achieved revenue exceeding ¥90B (+26.9% YoY). China has been the world’s largest industrial robot market for 13 consecutive years, with domestic brand market share exceeding 50%.

Sources: Science and Technology Weekly, MIIT

---

7 | LimX Dynamics (逐际动力) Closes ~$200M Pre-IPO at ¥15B Valuation

LimX Dynamics completed a ~$200M Pre-IPO round, bringing cumulative H1 funding above $400M and post-money valuation to ¥15B (~$2.1B). Funds earmarked for humanoid robot iteration, core component mass production, and commercial deployment. Company is now in formal IPO sprint phase.

Sources: Rongzhong Finance, 36Kr

---

8 | ZeroErr Cloud Control (零差云控) Closes Multi-Hundred-Million RMB C++ Round

Leading humanoid robot joint supplier ZeroErr completed a C++ round of several hundred million RMB led by Tongchuang Weiye. 2025 revenue doubled YoY, with humanoid/embodied AI orders contributing ~65% of growth. CEO notes the industry’s core challenge has shifted from “technology validation” to “supply chain capacity and consistency delivery.”

Sources: 36Kr/Hard氪

---

Track 4: Autonomous Driving (Robotaxi / L4 / Chips)

1 | XPeng Runs Full Robotaxi Chain in 8 Months — 3,000 TOPS, World’s Highest On-Vehicle Compute

XPeng CEO He Xiaopeng personally completed the first Robotaxi test ride on July 9 — from order to delivery, fully seamless. Based on the GX model with a pure-vision L4 solution, powered by 4 proprietary Turing AI chips delivering 3,000 TOPS (world’s highest on-vehicle compute). Running second-generation VLA large model. From announcement to internal testing: just 8 months. Q3 will launch demo passenger service in Guangzhou via Gaode Maps; 2027 will introduce economy models for driverless operation.

Sources: Road to Autonomy, XPeng Official

---

2 | Momenta Lists on HKEX: 414x Oversubscribed but Closes Flat on Day One

Momenta’s public offering was 413.6x oversubscribed with international orders exceeding HK$100B. 14 cornerstone investors including GIC and Fidelity ($100M each), plus BYD and Mercedes-Benz as strategic backers. Market share: 64.5%. But listing day told a different story: opened +6%, peaked at HK$314.8, then fell back to the HK$295.6 IPO price — zero gain. P/S ratio above 25x, net loss ¥303M, R&D at 77.5% of revenue. Secondary market is now pricing on financials, not narratives.

Sources: Road to Autonomy

---

3 | Global First: UN Approves Unified Autonomous Driving System Technical Regulation (ADS GTR)

The UNECE World Forum approved the first global ADS technical regulation, co-led by China, EU, US, UK, Canada, and Japan. Key provision: “ADS safety level must be at least equivalent to a competent and careful human driver.” China is simultaneously advancing a mandatory national standard (draft for July 2027 implementation). China’s L3 market penetration for ADAS-equipped passenger cars has reached 70%, with NOA-capable models exceeding 30%.

Sources: Xinhua, Economic Daily

---

4 | New Jersey Bill Would Ban Pure-Vision Autonomous Vehicles — Tesla Directly Targeted

New Jersey Senator Andrew Zwicker proposed S1677, requiring L4 commercial vehicles to have cameras plus at least two additional sensor types (lidar + radar), with a 3-year pilot, 50K-mile supervised testing requirement, and mandatory accident reporting. New York is considering similar standards. If passed, New Jersey would be the first US state to legislate autonomous vehicle hardware requirements — effectively banning Tesla’s Cybercab (no lidar, no steering wheel).

Sources: The Verge, Global Times

---

5 | Baidu Apollo Go: 22M+ Cumulative Orders, 27 Cities, 330M+ Autonomous Kilometers

Baidu’s Apollo Go has accumulated over 22 million orders across 27 cities globally, with autonomous driving mileage exceeding 330 million km (220M+ km fully driverless). The Wuhan operation covers a 3,000 sq km zone serving 7.7M residents including highway segments, and disclosed unit profitability in 2025.

Sources: Xinhua, Road to Autonomy

---

6 | NHTSA Clears Regulatory Path for Steering-Wheel-Free Vehicles

NHTSA has launched a three-step deregulation: eliminated mandatory transmission requirements, initiated revision of mandatory brake pedal requirements, and signaled steering wheel mandates will follow. The clear objective: remove federal hardware barriers for fully driverless vehicles like Tesla’s Cybercab.

Sources: Road to Autonomy

---

Track 5: World Models / Physical AI (Video Generation / Simulation / Compute Infrastructure)

1 | Huawei Atlas 950 SuperPod to Debut at WAIC — 8,192 NPUs in One System

Huawei confirmed the first-ever public display of Atlas 950 SuperPod at WAIC 2026 (July 17-20):

This represents the leap from single-chip to system-level domestic AI compute — positioned as surpassing NVIDIA’s NVL144 in interconnect bandwidth. WAIC stats: 1,100+ exhibitors, 100K+ sqm, 300+ global firsts.

Sources: Huawei Enterprise, WAIC Official, The Paper

---

2 | Sugon 8000: China’s First 100,000-Card All-Domestic AI Supercluster Goes Live

Sugon (中科曙光) announced Sugon 8000 (Dengfeng) — China’s first fully domestic 100,000-card AI supercluster — is now operational and connected to the National Supercomputing Internet. This marks the transition from 10K-card to 100K-card deployment scale for AI infrastructure, signaling that domestic compute capability is entering a new phase.

Sources: Cailian Press, Sugon Official

---

3 | Greater Bay Area Launches 10,000-Card Domestic Compute Cluster in Shaoguan

The Greater Bay Area domestic compute super-cluster went live in Shaoguan with 11,520 Ascend AI accelerator cards, total compute of 9,000P, and a total investment of ¥5.5B. The cluster serves GBA AI industry and marks the formal entry of domestic model R&D into a “compute self-reliance” phase.

Sources: Science and Technology Weekly

---

4 | Four Ministries Mandate 20 “AI + HRSS” Application Scenarios by Year-End

Four ministries (HRSS, NDRC, MIIT, National Data Administration) jointly issued implementation guidelines for “AI + Human Resources and Social Security”, targeting 20 application scenarios by end of 2026 across employment matching, social security processing, skills certification, and labor dispute resolution. A 67-scenario panorama map has been published.

Sources: CCTV News, Government Bulletin

---

5 | StepFun to Unveil World’s First “AI Agent Phone” at WAIC

StepFun (阶跃星辰), one of Shanghai’s “AI Six Dragons,” confirmed it will unveil the world’s first native AI agent smartphone at WAIC. Manufacturing partner: Huaqin Technology. Core feature: system-level Agent engine using GUI-based agent technology that can “see and control” any screen element across apps. Shareholders include Huaqin, Longcheer, OmniVision, and ZTE. StepFun has raised nearly $2.5B pre-IPO and is reportedly targeting a Hong Kong listing.

Sources: CNMO, Kuai Keji, Zhidongxi

---

6 | Kimi Teams with American Express and ABC for First “AI-Native” Credit Card

Moonshot AI’s Kimi partnered with American Express and Agricultural Bank of China to launch the first AI-native credit card — integrating AI compute credits as a card benefit. Cardholders can convert spending points into Kimi membership and token allowances. Future plans include agent-based payments, token-credit limit linkage, and points-compute bidirectional conversion.

Sources: Science and Technology Weekly

---

China AI Market Snapshot

---

Upcoming Events

---

Editor’s Take

The HunYuan Hy3 server crash is the most telling data point of the week.

When a 295B-parameter model from China’s largest internet company launches free and immediately overwhelms its own infrastructure — with 50% queue rates at peak — it tells you everything about the demand side of China’s AI equation. This isn’t hype; it’s genuine, overwhelming enterprise and consumer pull.

The MiniMax story is equally revealing: HK$16B raised at 7x oversubscription, a CEO taking zero salary until AGI, and a secret 2.7-trillion-parameter model in the works. Combined with DeepSeek V4’s groundbreaking peak-valley pricing (a concept borrowed from electricity markets — treating compute like a utility), China’s foundation model companies are moving from “catch up” to “set the rules.”

On the hardware side, the Unitree IPO is the clearest signal yet that humanoid robotics has crossed from venture-funded R&D into publicly-traded, profit-generating business. ¥600M non-recurring net profit, 60% gross margins, 5,500 units shipped in 2025 — this is a company that makes money selling robots today. AgiBot’s 15,000th unit and MIIT’s 100,000-unit annual forecast complete the picture: 2026 is genuinely the year of humanoid robot mass production.

XPeng’s 8-month sprint from Robotaxi announcement to live test rides, with 3,000 TOPS of on-vehicle compute, demonstrates the speed advantage of China’s integrated hardware-software approach. And Momenta’s 414x oversubscription followed by a flat close on day one is the market’s verdict: the scarcity premium is real, but the fundamentals must follow.

At WAIC next week, Huawei’s Atlas 950 SuperPod (8,192 NPUs, 16.3 PB/s interconnect) will be the centerpiece. If the specs hold under scrutiny, it’s the clearest signal that China’s domestic compute infrastructure is closing the gap — not at the chip level, but at the system level, where it matters most.

One-line summary: Server-crashing models, profit-making robot companies, peak-valley compute pricing, and 8,192-NPU supercomputers — China’s AI stack is going vertical, and it’s happening faster than anyone expected.

---

东声 East Voice — Physix Frontier’s China AI column for global readers
东声西达，中国视角。
© 2026 Physix Frontier · bbs.physixfrontier.com