🇨🇳 China AI Daily | 2026-07-13

July 13, 2026 (Monday) ｜ Team B · China → Global

Physix Frontier · bbs.physixfrontier.com

---

Editor’s Note

Today’s theme: “China’s AI ecosystem goes full-stack, open-source, and regulated — all at once.” Meituan drops LongCat-2.0, the first trillion-parameter model trained and deployed entirely on domestic chips, then open-sources four more projects in a single week. Chinese banks reveal jaw-dropping token consumption — 140 trillion tokens daily nationwide, with a single bank burning 33 billion tokens per day. Financial regulators respond with China’s first systematic AI governance framework for banking and insurance. Meanwhile, investors in AI startup Manus are reportedly trying to unwind its $2B Meta acquisition — a landmark case in AI sovereignty. And WAIC 2026 is four days away, with Huawei’s Atlas 950 SuperPod, StepFun’s AI agent phone, and over 300 global firsts on the way.

---

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

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

Meituan officially open-sourced LongCat-2.0, a 1.6 trillion parameter model with ~48B average active parameters — and the industry’s first trillion-parameter model to complete its entire training and inference lifecycle on a 50,000-card domestic computing cluster.

Key technical innovations:

• LongCat Sparse Attention (LSA): Reduces computational overhead from quadratic to linear for million-token contexts
• N-gram Embedding: A novel parameter expansion path beyond MoE experts, allocating 135B parameters to token-level representation (10% of total params)
• MOPD Architecture: Multi-teacher online distillation splitting experts into Agent, Reasoning, and Interaction categories
• Native 1M token context window, designed for Agentic Coding tasks

Benchmark highlights:

• SWE-bench Pro: 59.5 (vs Gemini 3.1 Pro 54.2, GPT-5.5 58.6, Claude Opus 4.6 57.3)
• SWE-bench Multilingual: 77.3 (vs Claude Opus 4.6 77.8)
• Terminal-Bench 2.1: 70.8

Available on HuggingFace, GitHub, and ModelScope in BF16, FP8, and INT8 variants, with optimized inference code for both GPU and domestic NPU (Huawei Ascend).

Sources: Meituan Tech Blog, 36Kr (Zhidx), AIToolly

---

2 | Meituan Launches “General 365” Benchmark: Even Gemini 3 Pro Only Scores 62.8%

Meituan’s LongCat team released General 365, a rigorous new reasoning benchmark designed to test LLMs beyond standard math and coding. In initial testing of 26 mainstream models:

• Gemini 3 Pro led with just 62.8% accuracy
• The vast majority of 26 models failed to reach 60%
• The benchmark is designed to expose weaknesses in complex logical reasoning, multi-step inference, and real-world problem-solving

This highlights that even frontier models still have significant room for improvement in robust reasoning — a critical capability for autonomous agents and enterprise applications.

Sources: GitHub Trending, Meituan LongCat

---

3 | Meituan Open-Sources VitaBench 2.0 and WBench: Two New Evaluation Frameworks for AI Agents

Meituan released two specialized evaluation benchmarks:

VitaBench 2.0 — the first benchmark for long-term dynamic user modeling in real-life scenarios. It systematically evaluates LLMs on two critical agent capabilities: personalization (adapting to individual users over time) and proactivity (anticipating user needs before they’re expressed). Unlike static task benchmarks, it simulates evolving, realistic interactions.

WBench — the first systematic multi-round benchmark for interactive video world models. Likened to a “CT scanner” for diagnosing AI limitations, it evaluates how models transition from passive observation to active interaction in dynamic environments — critical for robotics and embodied AI.

Sources: Meituan LongCat, AIToolly

---

4 | Meituan Open-Sources LongCat-Video-Avatar 1.5 and AIGC Poster Generation System

Two additional open-source releases from Meituan this week:

LongCat-Video-Avatar 1.5 — a digital human video model transitioning from academic SOTA to commercial-grade deployment. Key improvements: lip-sync accuracy, physical plausibility, long-video stability, multi-person interaction support, and optimized inference efficiency.

AIGC Poster Generation System — an industrial-grade visual content creation framework built around a unique “Generation-Editing-Evaluation” closed loop. Already deployed in Meituan Waimai and brand IP scenarios.

This week, Meituan has open-sourced five major projects — arguably the most prolific single-week open-source contribution from any Chinese tech company in 2026.

Sources: Meituan Tech Blog, AIToolly

---

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

1 | Chinese Banks Burn 140 Trillion Tokens Per Day — AI Spending Under Scrutiny as ROI Becomes the New Metric

Multiple Chinese banks disclosed token consumption data revealing the staggering scale of AI deployment in China’s financial sector:

National data from the National Data Administration: daily token consumption in China surged from 100 billion in early 2024 to 140 trillion by March 2026 — a 1,000x increase in two years.

However, banks are now pivoting from “burn tokens to show progress” to mandatory ROI assessment. CMB restricts coding-related AI to just 5% of compute (citing “spaghetti code” and security risks). Some banks are cutting minimum token consumption KPIs by 20% after executives grew alarmed at costs. One mid-sized bank estimates H2 token spending will reach ¥18M — 15% of its IT budget.

Sources: Economic Daily (经济日报), Securities Daily, EEO (经济观察网)

---

2 | China’s First Systemic AI Governance Framework for Banking and Insurance: 32 Rules Across 7 Dimensions

China’s National Financial Regulatory Administration (NFRA) issued the “Guiding Opinions on Safe Development and Application of AI in Banking and Insurance” — the country’s first comprehensive AI regulation for the financial sector. Key provisions:

Four core principles:

1. “Whoever uses it is responsible” — financial institutions bear full accountability
2. Autonomous and controllable — critical AI platforms must be self-developed
3. Pragmatic and efficient — no “AI for AI’s sake”
4. Safety-first development — cybersecurity and data security compliance mandatory

Critical requirements:

• High-risk AI applications (credit approval, claims, trading) require risk management committee approval
• Human oversight mandatory at key decision points for high-risk scenarios
• Personal data prohibited from being used to train generative AI models
• AI-generated content must be clearly labeled to consumers
• Algorithmic transparency and explainability standards required for high-risk use
• External AI models must pass internal security assessment and government filing

Separately, NFRA published draft Cybersecurity Management Rules for banking and insurance (8 chapters, 72 articles), with comment period open until August 10.

Sources: NFRA Official, Legal Daily (法治网), China News Service

---

3 | Manus Investors Seek to Unwind $2B Meta Acquisition at Same Valuation — A Landmark AI Sovereignty Case

According to Reuters, investors in AI agent startup Manus — including Tencent, ZhenFund, and Sequoia China — along with company management are discussing a deal to reverse Meta’s $2 billion acquisition at the same valuation.

Background: In April 2026, China’s National Development and Reform Commission (NDRC) blocked the Meta-Manus deal under the Foreign Investment Security Review mechanism — the first time China has vetoed an AI-related foreign acquisition on national security grounds.

Manus, which rose to fame in March 2025 as a breakthrough AI agent, relocated its headquarters to Singapore and shut down all Chinese operations after receiving US investment — a move widely criticized as “abandoning its Chinese roots.” The reversal, if completed, would mark a watershed moment in global AI sovereignty and cross-border M&A regulation.

Sources: Reuters, 36Kr, NDRC

---

4 | StepFun to Launch World’s First “AI Agent Phone” at WAIC — Ahead of OpenAI’s AI Device Timeline

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 2026 (July 17). Key details:

• Manufacturing partner: Huaqin Technology (华勤技术), a leading ODM
• Core feature: System-level Agent engine using GUI-based agent technology that can “see and control” any screen element, crossing app boundaries for seamless multi-step operations
• Example capability: “Order me the same coffee I had yesterday afternoon” — the phone autonomously opens the delivery app, finds the order, selects specs, and completes payment
• Shareholders include Huaqin, Longcheer, OmniVision, and ZTE — a deliberate model-hardware integration strategy

Meanwhile, Honor will showcase its Robot Phone concept at WAIC, and ZTE Nubia x ByteDance Doubao AI phone (second generation) is rumored to be a core WAIC launch. StepFun has raised nearly $2.5 billion pre-IPO and is reportedly targeting a Hong Kong listing as the third major model company to go public after Zhipu and MiniMax.

Sources: CNMO, Kuai Keji, Zhidongxi, Toutiao

---

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

1 | 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 — growing from just 6 prototype units in 2023 to 10,000 in March 2026 and 15,000 by June. The 5,000-unit increment took less than 3 months.

• 90% of components sourced from the Yangtze River Delta supply chain (some core parts available within hours)
• Robot models cover industrial handling, logistics sorting, and security patrols
• “Immediate delivery and deployment upon leaving the factory” is now standard operating procedure

Shanghai port exported 8.36 billion RMB (~$1.2B) in robots in the first 5 months of 2026, shipping to 113 countries and accounting for over 40% of China’s total robot exports.

Sources: Xinhua, China Business News

---

2 | Unitree Robotics IPO Approved: 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 in the history of the pre-review mechanism.

• IPO raises 4.2 billion RMB (~$580M) at an initial valuation of at least 42 billion RMB (~$5.8B)
• 2025 shipments: 5,500 units (global #1 for humanoid robots)
• Revenue: 1.708 billion RMB
• IPO proceeds: robot AI model R&D, new smart manufacturing base, global expansion
• Morgan Stanley raised its 2026 China humanoid robot shipment forecast from 28,000 to 50,000 units

Additionally, 51 robotics/embodied AI companies are currently in Hong Kong’s IPO queue (12.7% of all pending listings), enabled by the 18C Special Technology Company rules. H1 2026 embodied AI private funding exceeded 34.5 billion RMB.

Sources: Securities Times, 36Kr, China Business News

---

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

1 | WAIC 2026 Preview: Huawei Atlas 950 SuperPod to Make Public Debut — 8,192 NPUs in One System

With WAIC 2026 opening in four days (July 17-20 in Shanghai), Huawei confirmed the first-ever public display of its Atlas 950 SuperPod:

This represents the leap from single-chip to system-level domestic AI compute capability — and is positioned as significantly surpassing NVIDIA’s NVL144 in interconnect bandwidth.

WAIC 2026 stats: 1,100+ exhibitors, 100,000+ sqm exhibition space, 300+ global firsts. SpeedSensing (速腾聚创) will showcase new sensing technology July 18-20.

Sources: Huawei Enterprise, WAIC Official, The Paper (澎湃新闻)

---

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

1 | MIIT Targets 90% Agent Adoption by 2030 — “AI + Internet” as National Priority

At the 2026 China Internet Conference, MIIT’s Information and Communications Management Bureau declared “AI + Internet” a core national development direction, with a hard target: 90% adoption rate for next-generation intelligent terminals and AI agents by 2030.

The National Development and Reform Commission revealed that China’s daily token consumption has surpassed 140 trillion (up from 100 billion in early 2024). Academician Wu Hequan projected that 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%
• Agent services reach 75% of AI traffic

China’s three telecom giants (China Mobile, China Telecom, China Unicom) announced strategic pivots toward “Token Economy” operations — treating tokens as a utility akin to electricity.

Sources: CGTN, NDRC, China Internet Conference

---

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

Four ministries including the Ministry of Human Resources and Social Security jointly issued an opinion 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 — from employment matching to social security processing to skills certification.

Sources: CSDN, Government Bulletin

---

China AI Market Snapshot

---

Upcoming Events

---

Editor’s Take

This week’s Meituan open-source blitz deserves more attention than it’s getting.

Five major projects in one week — a trillion-parameter model (LongCat-2.0), two evaluation benchmarks (VitaBench 2.0, WBench), a digital human model (Avatar 1.5), and an AIGC poster system — isn’t just prolific output. It’s a strategic statement: Meituan is positioning itself as an AI infrastructure provider, not just a food delivery company.

The real story in LongCat-2.0 isn’t the parameter count. It’s the fact that it was trained and deployed entirely on domestic chips — 50,000 cards of homegrown compute, with optimized inference code for both GPUs and Huawei Ascend NPUs. When the Alibaba Claude ban forced 120,000 developers onto domestic tools last week, LongCat-2.0 arrived as proof that the domestic stack can handle trillion-parameter workloads. This is the “sovereign AI” thesis made real.

Meanwhile, the banking token consumption data is staggering in scale but revealing in its backlash. When China Merchants Bank — the industry leader — says coding AI produces “spaghetti code” and restricts it to 5% of compute, that’s a cautionary tale for every enterprise rushing to deploy AI coding assistants. And when mid-sized banks find that token spending is eating 15% of their IT budget with questionable ROI, the industry’s pivot from “burn more tokens” to “prove the ROI” is both rational and overdue.

The Manus-Meta unwinding is perhaps the most geopolitically significant story of the week. China’s first-ever veto of a foreign AI acquisition on national security grounds — and investors now trying to reverse it at the same valuation — sets a precedent that every global AI company must factor into their M&A strategy.

In four days, WAIC opens with Huawei’s Atlas 950 SuperPod as the centerpiece. 8,192 NPUs, 16.3 PB/s interconnect, fully liquid-cooled. If the specs hold under scrutiny, it’s the clearest signal yet that China’s domestic compute infrastructure is closing the gap — not at the chip level, but at the system level, where it matters most for training and deploying frontier models.

One-line summary: Open-source models on domestic chips, token-economy banking, AI sovereignty vetoes, and 8,192-NPU supercomputers — China’s AI stack is going vertical.

---

China AI Daily is produced by Physix Frontier’s East Voice editorial team.
© 2026 Physix Frontier · bbs.physixfrontier.com