|
|
|
|
Bottom Line Up Front 360 Gruop unveils Tulong Feng at ISC.AI 2026 in Beijing on June 24, with founder Zhou Hongyi positioning the vulnerability discovery agent as a domestic answer to Mythos. The model reportedly discovered 3,432 vulnerabilities with 105 confirmed by regulators. MIIT codifies new five-year tech sequence at MWC Shanghai with chief engineer Zhong Zhihong naming AI, quantum technology, embodied intelligence and 6G as priority innovation fields. Huawei to debut Atlas 950 datacenter AI training system at WAIC Shanghai from July 17 to 20, the first showing since the March 2026 launch of the system. Atlas 950 reportedly links up to 8,192 Ascend NPUs at 64 per rack to enable trillion-parameter AI model training runs. |
| 360 unveils Tulong Feng as China's domestic counter to Anthropic's Mythos AI vulnerability discovery model At the ISC.AI 2026 cyber conference opening in Beijing on June 24, 360 Group founder Zhou Hongyi launched two AI security capabilities: vulnerability discovery agent Tulong Feng and automated defense system Yitian Zhen. He said Tulong Feng has already surfaced 3,432 vulnerabilities, with 105 confirmed by regulators and several rated high-severity in the national vulnerability database. 360 also launched a Rock Shield security coalition with Phytium and Hygon (domestic CPUs), Kylin (operating system) and others, modeled on what Zhou described as a U.S. Glasswing alliance built around Mythos for U.S. critical infrastructure. |
| MIIT chief engineer at MWC Shanghai codifies dual-ten-gigabit, 6G and embodied AI as the next five-year tech sequence At the MWC Shanghai (Mobile World Congress) opening on June 24, Ministry of Industry and Information Technology (MIIT) chief engineer Zhong Zhihong called for "maintaining measured forward-positioning" by pushing China's dual-gigabit (gigabit-per-second fixed and mobile) networks toward dual-ten-gigabit service and building a multi-layer compute infrastructure paired with low-altitude information infrastructure and satellite internet. Zhong named artificial intelligence, quantum technology and embodied intelligence (AI driven humanoid and physical systems) as priority innovation fields alongside 6G core technology R&D and standards work. MIIT licenses telecom and sets industrial policy in China. |
| Huawei will debut the 8,192-NPU Atlas 950 SuperPoD physical system at WAIC Shanghai in July Huawei said on June 24 it will physically demonstrate its Atlas 950 SuperPoD at the 2026 World Artificial Intelligence Conference (WAIC) in Shanghai from July 17 to 20, the first public showing of the system since its March 2026 launch. The Atlas 950 packs 64 Ascend NPUs (neural processing units, Huawei's domestic AI accelerators) per rack and supports up to 8,192 chips interconnected as a single training fabric for trillion-parameter model training and inference. Huawei will also display the Atlas 850E air-cooled variant with an upgraded VCE thermal design and continue open-sourcing the Ascend AI software stack. |
| Premier Li Qiang at Summer Davos says China's "door of opening will only open wider" At the opening of the 17th World Economic Forum’s Annual Meeting in Dalian on June 24, also referred to as the 2026 Summer Davos forum, State Council Premier Li Qiang said "the door of China's opening will only open wider" regardless of "international winds and rain" and urged participants to "tear down walls, build bridges" through global innovation cooperation. Li reiterated that Q1 GDP growth came in at 5% and said Q2 has maintained "good operational momentum," casting the country's 140-trillion-yuan ($19 trillion) economy as a source of "precious certainty" and a "safe haven" against international energy shortages and supply chain disruption. State broadcaster CGTN had previewed earlier this week that Li would attend the opening plenary and deliver a special address. The forum's theme this year is "Innovating at Scale." |
|
| NRTA draft rules require labels on AI generated micro-dramas China's National Radio and Television Administration (NRTA), which licenses television and online video content, published draft Micro-Drama Development Management Measures for public comment on June 24, requiring AI generated or AI assisted micro-dramas to display visible prompt labels on every episode. The rules sort micro-dramas (short serialized episodes under 20 minutes designed for phones) into three tiers by investment scale and subject matter, with Tier 1 covering large-budget productions or any treatment of politics, military, national security and other officially sensitive topics. Tier 1 productions require pre-filing review and a distribution license before broadcast, while Tiers 2 and 3 use a lighter classified-filing-and-disclosure regime. |
| WEF adds 8 China sites to its Lighthouse Factories list, pushing the China total to 109 and 45% of the global network The World Economic Forum announced 16 new Lighthouse Factories at the Dalian Summer Davos on June 23, with 8 of the new sites located in China including NIO's Hefei electric vehicle plant, Sinopharm Taiji's Fuling pharmaceutical plant and Haier's Rrs supply chain park in Qingdao. Wenweipo reports that China's total Lighthouse count has now reached 109 of 238 globally, the largest single-country share at roughly 45%. A Lighthouse Factory is part of the World Economic Forum's Global Lighthouse Network, which recognizes manufacturing sites that deploy advanced technologies like AI and robotics at scale. |
|
| III | AI & FOUNDATION MODELS | |
| ByteDance's Doubao 2.1 Pro claims parity with Claude Opus 4.7 on coding benchmarks At ByteDance's June 23 FORCE conference, Volcano Engine released Doubao-Seed-2.1-Pro and demonstrated the model running unattended for nearly 18 hours through 9 iterations on a 16-by-16 PE Tiny NPU Tile, producing 6 core modules and 1,303 lines of register-transfer-level code (the digital circuit description language chip designers normally hand-write over weeks). On Terminal Bench 2.1, a coding benchmark simulating production developer environments, ByteDance reports Doubao 2.1 Pro tied with Anthropic's Claude Opus 4.7, and on the SciCode and MCP-Atlas evaluations it reported scores above both Opus 4.7 and OpenAI's GPT-5.5. ByteDance priced Doubao 2.1 Pro at 6 yuan ($0.83) per million input tokens and 30 yuan ($4.15) per million output tokens, roughly one-quarter of the price of Opus 4.6 and 4.8. |
| Baidu open-sources Unlimited-OCR, solving bottlenecks in long-document transcription Baidu open-sourced Unlimited-OCR on June 22 via GitHub and HuggingFace, a Mixture-of-Experts model with 3 billion total parameters and only 500 million active during inference (the working subset used per query), per AI outlet Xinzhiyuan carried by iFeng Tech. The model scored 93.23% on OmniDocBench v1.5 and 93.92% on v1.6 (both end-to-end document parsing benchmarks where higher is better), ahead of Alibaba's 235-billion-parameter Qwen3-VL at 89.15 and Google's Gemini 2.5 Pro at 88.03. Its Reference Sliding Window Attention mechanism holds the inference cache at fixed size, letting the model transcribe 40-page documents with a 0.11 edit-distance error against ground truth in a single inference pass per Baidu's reported tests. |
| Tesla mainland fleet to feature ByteDance's Doubao voice system for vehicle control, with DeepSeek handling chat At ByteDance's June 23 FORCE conference in Beijing, Tesla China confirmed its mainland fleet will integrate Doubao for vehicle control voice commands (climate, windows, navigation) and DeepSeek for general conversational queries, per ITHome. Drivers in mainland China will wake the assistant with "Hey, Tesla" or the steering wheel voice button, and Tesla's internal software build 2026.14.11 already shows a Doubao AI assistant entry in the application list. ByteDance and Tesla had publicly named Doubao and DeepSeek Chat as the planned local voice providers in Tesla's August 2025 owner-terms update, which the carrier described at the time as pending a rollout date. |
|
| ByteDance in talks for $20 billion offshore loan, its largest ever, to fund AI and cloud computing expansion ByteDance is in preliminary discussions with multiple banks for an offshore loan of about $20 billion on a three-year tenor extendable to five years, per Bloomberg via ITHome. It would be the company's largest offshore borrowing to date and is expected to fund AI compute and cloud expansion. |
| Daxiao Robotics closes Angel+ round with national, automaker and chip capital around its open-source Kairos world model Humanoid robotics startup Daxiao Robotics, announced June 23 that it closed an Angel+ round with Shenzhen Capital Group, Geely Capital and Metax (the GPU designer) among the named investors, with existing backer SenseTime Guoxiang Capital increasing its stake. Per Leiphone, Daxiao has cumulatively raised "hundreds of millions of dollars" across two rounds inside four months. Daxiao also open-sourced Kairos world model 3.0, which the company says ranked first on four embodied AI benchmarks including RoboTwin 2.0 at 96.1 and LIBERO-Plus at 89.0, ahead of NVIDIA's Cosmos 3 and Alibaba's Wan 2.2 on Daxiao's reported scores. |
| Striding AI closes nearly $100 million Angel round with CP Group Physical AI startup Striding AI announced on June 23 that it has closed an angel round of nearly $100 million from backers including Thailand's CP Group, Chinese listed manufacturer Huaqin Technology, medical device maker Andon Health, and several prominent entrepreneurs and top-tier investment firms. Striding AI said the proceeds will fund hiring of top global algorithm, engineering and commercialization talent, iteration of core technologies including its world action model and reinforcement learning work, and rollouts in retail and industrial settings. The company said it is building on deep partnerships with CP Group and Huaqin to develop real commercial and industrial deployment scenarios |
|
|
|