Alibaba Tongyi Qianwen releases multimodal agent model Qwen3.7-Plus, integrating vision, language and Agent capabilities, which can understand images and complete tasks manually. It is now online on Alibaba Cloud Bailian platform, with online experience open to users.
MuleRun officially launches Messages, an AI collaboration IM for MuleRun Enterprise edition. It supports human-human, human-Agent and Agent-Agent collaboration in the same workspace, reduces information transfer and improves organizational collaboration efficiency. MuleRun Enterprise users can start trying it now.
ByteDance's Seed adjusts its organizational structure: the original Seed Robotics team led by Li Hang has been reporting to Zhou Chang, head of multimodality, for over a month. Li Hang now serves as an advisor for academic cooperation. ByteDance is recruiting a technical head of embodied AI who will report to Zhou Chang.
Embodied intelligence company Xinghaitu has released a bipedal robot "Kengo". Standing 1.4 meters tall, its single joint torque exceeds 130N·m, capable of high-dynamic motions. It can still operate normally after falling 10 times, with a wiring harness bending life of over 200,000 times. This is Xinghaitu's first bipedal robot product.
In AGI Ranker's latest coding ranking, Anthropic's Claude Opus 4.8 topped the list with 81.01 points, 3.5 points ahead of OpenAI's GPT-5.5. Anthropic has confidentially submitted an IPO application, OpenAI plans to release a major Codex update, and the market predicts GPT-5.6 will likely be released this week.
Apple will unveil iOS 27 at WWDC26, with a focus on upgrading AI capabilities, enhancing Siri, adding AI features to multiple system apps, supporting on-device AI execution, and supporting iPhone 12 and newer models.
Gu Quanquan, core member of the AI4S team at ByteDance's Seed, announced his departure. During his time at ByteDance, he developed multiple AI models for biomolecules and proteins with leading performance, and may start a business in the AI4S field.
Huxiu held a closed-door meeting on AI hardware, where four industry professionals conducted in-depth discussions on AI hardware development, pointing out that industrial AI has achieved earlier implementation, early sales of consumer robots mostly come from intra-industry circulation, and emphasizing the importance of PMF and computing power adaptation.
AhaCreator, the world's largest AI native influencer marketing platform, has launched version 3.0. In 1 year since launch, it has attracted 100,000 creators covering 140+ countries, served over 300 brands, and fully automates the entire influencer marketing workflow to significantly improve operational efficiency.
Institute of Automation, Chinese Academy of Sciences open-sources MobileGym, an Android simulation platform running in browsers, used for training and evaluating mobile AI Agents, supports 28 daily apps, each instance takes up only ~400MB memory, with much lower cost than emulators and real devices.
Inspired by a post on X, a developer built an open-source computer cleaning skill based on Codex. It supports Mac/Windows, can analyze storage by risk levels and assist cleaning. The developer cleaned up ~120GB of space in testing, and it's now free and open-source on GitHub.
After posting a net loss of 176 million yuan in 2025, iReader has launched the 999-yuan AI trendy toy CreMoMo for pre-order. It is equipped with a large emotional model for personalized companionship, and has a built-in AI Agent that supports tasks such as ticket grabbing. As of writing, 19 units have been sold on Taobao and 5 on Douyin.
On June 1 local time, Google's parent company Alphabet announced it plans to raise $80 billion to expand AI infrastructure. Berkshire Hathaway will invest $10 billion. This year Alphabet has raised funds through multiple channels to expand AI investment.
Deep Principle team has produced Materials Property Axiom (MPA), a material foundation model, through recursive self-training using its self-developed MIRA AI scientific research platform. MPA sets new SOTA across 40 experimental property prediction tasks, with average error reduced by 10% and maximum reduction reaching 51%.
Author Ai Doubi migrated his OpenClaw Agent hosting service to FastClaw framework based on compute-storage decoupling architecture, reducing servers from 18 to 3, operating cost to 1/6 of original. FastClaw has 1/40 code volume and 1/7 resource usage of OpenClaw.
A Reddit user developed the ADHD Skill tool, which enables Claude Code to think in an ADHD-style. Through multiple independent Agents, multi-perspective division of labor, and separation of generation and criticism, it improves the novelty and breadth of solutions. Tests show novelty increased by 5.17 and breadth increased by 4.17, installable with one command.
Singer Hu Yanbin developed the fan interaction app "Yan Huo" using Trae tool via vibe coding, which uses ByteDance's Doubao (Lark) model. The app has been found to have multiple security vulnerabilities, including unmoderated user content, unprotected SMS interface, etc.
Chinese AI chip company MooreX has completed Series C financing of nearly 1 billion yuan. The company uses sparse computing technology to reduce AI inference costs, and its new generation general-purpose AI inference computing card SparsePrime® will be launched within this year.
Shanghai Jiao Tong University and other universities/institutions release DataMaster, which enables AI to automatically complete data searching, filtering, cleaning and other work under fixed models and algorithms, and continuously iterates to optimize data. Experiments show it effectively improves model performance.
In May 2026, Zhixiang Future open-sourced the 8B-parameter text-to-image model HiDream-O1-Image, which topped the global ranking of open-source text-to-image models on Artificial Analysis with an Elo score of 1187. It adopts the innovative UiT architecture to reduce information loss and supports multiple tasks, but its ecosystem construction is still in its initial stage.
Blogger "Laojin Takes You Play AI" shares a usage method of Codex with Coze 3.0 and Claude Code, enabling remote control of multi-threaded AI automatic work via mobile phone, with final results reviewed by Claude Code to achieve laid-back development.
Turing Award winner Richard S. Sutton and co-author published a paper, criticizing the passive representation approach of current AI that relies on static data, introducing enactive cognition from cognitive science to AI, advocating that AI forms cognition through interaction with the environment.
In 2026, multiple companies have withdrawn their full AI adoption plans due to high AI Token costs: Microsoft canceled some Claude Code licenses, Klarna rehired human customer service, Duolingo adjusted AI performance policies, and one enterprise spent over $500 million on Tokens in a single month.
In May 2026, Pope Leo XIV partnered with Anthropic co-founder Christopher Olah to warn that AGI will arrive in 2030, leaving humanity only about three years to establish a unified global regulatory framework for self-rescue. Scholars proposed adopting a global financial regulation model to govern AI.
CI-Yuan Wu-Xian (Tokfinity), founded by a team from ByteDance and Tsinghua Yao Class, has completed tens of millions of RMB in Series Angel+ financing. Within 6 months of establishment, it has secured multiple multi-million RMB orders, and is committed to defining the global enterprise-level Agent infrastructure. The funds will be used for product commercial expansion.
Mind Lab (Mindverse) has consecutively released multiple research advances in the field of LoRA and PEFT, proposing a new paradigm for continual learning of large models, covering δ-mem memory mechanism, MinT infrastructure, etc., which can support millions of small agents running on large base models.
HeyGen has launched the open-source product HyperFrames, featuring the Vibe Motion solution that generates videos through AI-written HTML code. It is positioned for AI agents rather than regular users, and HeyGen's current ARR exceeds $100 million.
Startup Xiangmuguo Robotics emerges from stealth mode, founded by 9 Tsinghua University PhDs. It proposes an 'instinct'-based robotic operation technical route that enables robot operation without training data. The company has completed nearly 100 million RMB seed round financing and entered industrial scene implementation.
After one year of tracking AI healthcare startup Facere in Australia backed by Impa Ventures, Facere develops AI-powered medical workflow scheduling and voice assistant products, the company's MRR has increased 12x. The article proposes three stages of AI healthcare development, pointing out that the core value of AI healthcare lies in restructuring medical workflows.
Large model company Zhipu (valued at approximately 564.1 billion yuan) plans to apply for A-share listing on the STAR Market, raising 15 billion yuan, with 80% of the funds to be used for AI general foundation model R&D. The plan has been approved by the board of directors and will be submitted to the general meeting of shareholders for approval.
Tsinghua University AIR, in collaboration with multiple institutions, launched UniLab, an open-source robotic reinforcement learning training architecture. It adopts CPU simulation + GPU training architecture, achieving 3-10x faster training speed, can train a humanoid robot in 3 minutes, supports cross-platform operation, and can run locally on Mac.
Li Yitong, former Huawei 'Genius Youth' and former head of large model technology at Huawei Terminal Cloud, has joined embodied intelligence company Jiyi Intelligence, serving as chief engineer of the large model R&D center and leading core technology breakthroughs. Founded in August 2025 with registered capital exceeding 1.6 billion yuan, Jiyi Intelligence will launch a global high-salary recruitment program.
MiniMax releases new generation large model M3 and supporting Agent product MiniMax Code. M3 features 1M context window, cutting-edge coding capability and native multimodality. Officially, its performance reaches the level of Sonnet 4.6, and model weights will be open-sourced within ten days.
StepFun officially released and open-sourced Step 3.7 Flash, a new generation Flash model for production-grade Agents. Its programming capability reaches 97% of Claude Opus 4.6 while task cost is only 1/9 of the latter. It supports multimodal understanding and high-throughput parallel inference, and is compatible with mainstream Agent tools.
China Merchants Group Lion Rock Artificial Intelligence Laboratory has developed self-developed LiOS operating system, connected the full链路 of embodied intelligence, realized cloud-edge collaboration, supported robots to complete clothing folding tasks in real scenarios, and will open source the low-latency image transmission module and LeFold dataset recently.
Aippy, the world's first AI game community, has completed tens of millions of dollars in Series A financing led by Glowill Capital, with a post-investment valuation of $250 million (approximately 1.7 billion yuan). Incubated by Chizi City Technology, it lowers the threshold for game creation and has exceeded 3 million global downloads so far.
A report by moises and Water & Music shows 78% of professional musicians already use AI, and over 44% of new uploads on Deezer are fully AI-generated. Major music platforms have launched authentication, labeling or detection measures, and the industry believes musicians will rely on unique personality to break through in the future.
Multiple OPC (one-person company) entrepreneurs in Haidian, Beijing, have launched startups at low cost with the help of AI, covering multiple fields. Suyuan Matrix secured nearly 30 million yuan in orders within one year of establishment. Haidian has introduced multiple policies to support OPC entrepreneurship, and entrepreneurs ignore doubts and choose to develop by leveraging AI.
OpenAI CEO Sam Altman publicly retracted his previous AI "apocalypse of jobs" at the Commonwealth Bank of Australia Sydney conference, admitting that AI is disrupting employment slower than expected, and human interactions are hard to be replaced by AI. Multiple tech leaders agree that AI job loss claims have been overstated.
Weimai has again submitted an IPO application to the Hong Kong Stock Exchange, sprinting to become the "first AI full-course disease management stock". As of May 2026, it has cooperated with 188 hospitals, achieved revenue of 863 million yuan and net loss of 290 million yuan in 2025, and must complete the IPO by the end of 2026 or the redemption clause will be triggered.
Researchers from University of Oxford, NVIDIA and other institutions jointly propose KV-CAT, a new training method that lets models learn to generate compressible KV cache during training, which can significantly improve subsequent compression performance and long text processing capabilities without losing basic model capabilities.
WeChat public account "markzou's Notes" released an interpretation of AI large model intelligent travel planning project plan, introducing project process, division of labor between large models and traditional programming, evaluation system, prompt design, model selection and tuning methods, solutions to common problems, and provides an H5 for trial use.
A team from Tianjin University and Shanghai Jiao Tong University proposed the SE-GA framework at ICML 2026. It solves the problems of existing GUI agents that cannot remember historical information and learn from experience through hierarchical memory and self-evolution mechanism, and achieves leading performance on multiple benchmarks.
Invasive brain-computer interface startup SiClink (XiLian Tech) completes tens of millions of RMB seed and angel rounds co-invested by BlueRun Ventures, Hillhouse Capital, etc., focusing on visual reconstruction and developing silicon-carbon interconnected bidirectional brain-computer interface.
China technology study tours have gained popularity overseas. The tours last 3-5 days, priced from thousands to tens of thousands of US dollars, covering visits to AI, robotics, electric vehicle and other technology companies. One institution has received over 1,000 foreign visitors in 18 months.
South Korean AI inference chip developer XCENA has completed a $135 million Series B funding round, reaching a valuation of $570 million with total cumulative funding of $185 million. The company develops the MX1 chip, which places computing near memory to reduce data transmission and lower AI infrastructure costs.
Anthropic has confidentially submitted an S-1 prospectus to the U.S. SEC ahead of OpenAI,冲击 what could be the largest AI IPO in history, with a valuation approaching $1 trillion, annualized revenue reaching $470 billion, and expects to go public as early as October this year.