Alipay, owned by Ant Group, launched the world's first Token Pay service and AI wallet, completing the construction of a full-stack AI-native payment system. Currently, AI Pay has processed 300 million agent payments, supporting 95% of general agent frameworks, becoming the world's first large-scale commercial AI-native payment infrastructure.
Chinese company MengXiaoi from Hangzhou launches PettiChat AI pet translation collar, priced at around 800 RMB, with 94.6% accuracy, powered by Qwen-Omni large model, has received 863 crowdfunding backers, raising 140,000 HKD.
The Shanghai Stock Exchange announced that Unitree Robotics' STAR Market IPO will be reviewed on June 1. It only took 66 days from acceptance to review, setting a new STAR Market review speed record. In 2025, its revenue reached 1.699 billion yuan, with humanoid robot business accounting for over 50%, and the initial market capitalization is expected to be at least 42 billion yuan.
"Father of Lobster" Peter open-sources the skill-cleaner tool, which can analyze and audit AI skills, streamline descriptions, detect duplicates, screen for unused skills, helping save tokens, reduce costs and improve AI skill calling accuracy.
Pixmax launches a one-stop AI short drama creation platform, aggregates multiple mainstream large models, supports multi-person collaborative creation, and centrally manages team personnel, assets and creation costs. Enterprise certification is free with no limit on the number of sub-accounts, and a redemption code is provided for 500 points.
The RLinf framework jointly launched by Wuwen Xinqiong and Tsinghua University has systematically reconstructed the training pipeline of NVIDIA's DreamZero world model, achieving nearly 4x training throughput improvement, reducing the original 25-day training to about one week while maintaining model convergence quality.
ByteDance-backed MiniMax (Mianbi Intelligence) launches ForgeTrain, the world's first production-level large model pre-training framework fully written by AI. Its performance exceeds NVIDIA's Megatron, with 10% training acceleration on Huawei Ascend. It also trains the small-size edge model MiniCPM5-1B, which is now open source.
AI startup Hark has completed $700 million Series A funding, with a post-investment valuation of $6 billion. Multiple tech companies including NVIDIA participated in the funding. Founded in late 2025, Hark plans to build a consumer-facing personal AI assistant platform and native AI hardware, with products still undisclosed.
Kunlun Wanwei launches domestic native Agent models SkyClaw-v1.0 and its lite version, with performance ranking among the world's first tier. The price is only half or even lower than mainstream top-tier models, currently available for free for a limited time, compatible with major Agent frameworks and OpenAI interfaces.
In May 2026, Kaiming He's team from MIT and ByteDance's Seed Lab almost simultaneously published papers, confirming that language generation can be completed in continuous space and only mapped back to text at the final step, achieving better, faster and more cost-effective results than traditional discrete token methods, opening up a new technical path for AI large models.
The MINT team from Shanghai Jiao Tong University launched Evo-Depth, a solution with only about 0.9B parameters that adds spatial perception capability to VLA without increasing hardware burden, achieving an average success rate of about 90% on real robots. Code and model weights have been open-sourced.
A user asked Doubao about airline ticket refund fees. Doubao said only 5% would be deducted and promised to compensate 600 yuan if the fee wasn't recovered. However, the actual fee was 40% and no compensation was made, so the user sued ByteDance. The incident exposes the promise hallucination problem of conversational AI, and currently AI promises are not legally regarded as the manufacturer's declaration of intent.
ByteDance plans to increase its spending on AI infrastructure by 25% to 200 billion yuan in 2026, focusing on core projects such as Doubao and cutting about 30% of non-core AI application projects. Currently, Doubao has 449 million monthly active users, ranking first in the industry, and is exploring paid commercialization.
A research team from Zhejiang University and Alibaba proposes Unified Thinker, a reasoning framework that decouples thinking from generation, improving logical reasoning and generation capabilities of open-source vision models. The work has been accepted as Oral at ACL 2026.
OpenAI co-founder and president Greg Brockman recapped the 72-hour crisis of Sam Altman's removal in 2023 for the first time on a podcast, shared the course of events, changes in the company's governance structure, and discussed his views on the future of AI R&D.
UniPat Laboratory launches SaaS-Bench, which evaluates AI Agents using real SaaS office environments and database verification. It covers 106 tasks across 6 fields, most are cross-App tasks over 100 steps. Tests show top models Opus and GPT-5.5 lead in performance, but even the highest score is less than half.
Researchers release Spatial-Agent, which can convert complex geographic problems into executable geographic analysis workflows, achieves significant accuracy improvement on multiple benchmarks, and is compatible with different large language models.
Three channels are available for free/low-cost access to DeepSeek models: Nous Portal provides DeepSeek V4 Flash for free; Freebuff offers 500 free monthly requests for DeepSeek V4 Pro; Reasonix reduces DeepSeek API costs through caching optimization.
Elon Musk's xAI has completed training of its Grok base model V9-Medium (1.5T), which incorporates large amounts of Cursor data to improve programming capabilities, expected to release in 2-3 weeks. Meanwhile, AI programming tool Grok Build has entered early Beta testing.
Altimeter Capital partner Jamin Ball disclosed that Anthropic plans to pay SpaceX nearly $45 billion over three years for computing power, with an annual fee of $15 billion and monthly fee of $1.25 billion, a deal that will nearly double SpaceX's annualized revenue.
With the improvement of AI multimodal generation capability and the reduction of AI deepfake cost, AI-generated fake news has occurred frequently. Agent Universe has launched an interactive AI image identification platform "ZunDuo JiaDuo" for users to train their ability to judge authenticity.
Coolpad Tech's spatial intelligence platform Aholo open-sources Aholo Viewer, a 3D Gaussian Splatting browser that outperforms Spark 2.0 from Fei-Fei Li's World Labs, enabling smooth loading of large 3D scenes with over 1 billion particles in any device browser, allowing users to browse 3D worlds just like scrolling videos.
One-person AI company Polsia raises $30 million in funding, reaching a valuation of $250 million. The platform can automatically develop and operate companies based on users' ideas, claims to reach nearly $10 million ARR six months after founding, and currently faces widespread controversy.
Sam Altman, CEO of OpenAI, offers each of the 169 startups in this Y Combinator batch $2 million worth of OpenAI tokens in exchange for company equity, with a total value exceeding $300 million. The tokens can be used for GPT model calls and other AI development.
OpenAI Codex team member Vaibhav Srivastav has publicly released a self-distillation prompt. After copying and pasting it into Codex, the AI can automatically identify users' repetitive work and package it into reusable tools. It has been iterated to version 2.0 covering multiple work scenarios, and OpenAI President Greg Brockman retweeted and liked it.
DeepSeek announces permanent price reduction for DeepSeek-V4-Pro model API to one quarter of the original price, completes output speed-up and service expansion, supports 500 concurrency by default, lowers developers' usage costs, and plans to build a Chinese version of Claude Code.
Kimi completes its Series C funding with a post-investment valuation of $4.3 billion, led by IDG with an investment of $500 million. Today Capital's Xu Xin promoted this financing; Xu Xin is an early investor in Kimi and has supported the company's development multiple times before.
The VeRL-Omni team releases a general RL post-training framework for diffusion and full-modal generative models, supporting multiple architectures, compatible with NVIDIA GPUs and Ascend NPUs. Separating reward model reduces per-step time by ~14%, code is open-sourced.
Ashpreet Bedi, founder of Agno, has launched the open-source project Scout, positioned as a "company brain" that integrates company knowledge scattered across various tools. It adopts a navigation-over-search approach to solve the problem of information access for AI Agent, and is now open source.
Alibaba's Accio Work launched an enterprise version, supporting team members to share Skills and Agents, with functions including member management, permission management, automatic updates, etc. The entry version supports up to 5-member teams, and 15-member and 50-member versions are also available.
Research teams from USTC, UCLA, CUHK and Xiaohongshu propose Flow-OPD, the first multi-task post-training framework introducing multi-teacher On-Policy Distillation (OPD) to flow matching models, which solves the trade-off problem in multi-task image generation training and outperforms single-domain expert models.
JD.com's JoyInside head Dai Wenjun introduced the "AI World" concept at AIGC Summit 2026, envisioning AI actively integrating into home devices like robots and toys, moving beyond screen-based interaction. Leveraging JoyAI foundation model and super supply chain, JD.com targets embedding AI into over 10 million devices by end of 2026, building the world's largest physical world operations center to bring AI from industries to households.
Mureka launches an AI music generation model based on the MusiCoT framework, enabling AI to plan before creating. Its output quality meets production-ready standards, with more than 8,000 enterprises worldwide having connected to it, and multiple clients have achieved significant business growth after integration.
U.S. biotech startup Bexorg has maintained more than 700 post-mortem donor human brains ex vivo and conducted drug tests, keeping the brains viable for approximately 24 hours to assist with neurological drug development and more accurately predict drug efficacy, and has raised $42.5 million in funding.
Google releases Gemini 3.5 Flash at Google I/O conference. Positioned in the Flash series, it achieves 83.6% SOTA on Agent benchmark, with output speed of 289 tokens/second and first token latency of 65 milliseconds, about 4x faster than other cutting-edge models. Priced at $1.5/1M input tokens and $9/1M output tokens.
Fengxing Online CEO Yi Zhengchao shared at the 2026 China AIGC Industry Summit that the company has spent three years training all employees (including non-technical positions) in AI coding, promoted corporate AI transformation, achieved more than 10 times profit growth, and is currently focusing on AI video and mass innovation ecosystem building.
American Hunter Bown's open-source project DeepSeek TUI (renamed Codewhale) has gained 34.5K Stars on GitHub. After coming to China for community exchanges, he claimed he seemed to be used and left China early, triggering controversy in the open-source community.
FDE (Forward Deployed Engineer), originally an internal jargon at Palantir, has now become a popular new job title in the AI era. OpenAI founded Deployment Company with an initial investment of $4 billion in May 2026, and Anthropic is also recruiting for this position on a large scale. This role refers to engineers stationing at customer sites to help implement AI models into customers' actual business operations.