Jina AI releases jina-embeddings-v5-omni, an omni-modal embedding model supporting text, image, audio and video. Only 0.35% of weights are trained, it matches performance of larger models with smaller parameters, and is compatible with existing text indexes without rebuilding.
Hu Debo, former CEO of Kepler Robotics, founded new company Sota Unbounded, betting on embodied intelligent robot brain. It plans to showcase full capabilities this summer, focusing on European and American retail scenarios for global expansion, and has secured a framework agreement with a leading European supermarket, with predicted order volume of nearly 1,000 units.
Google launched Magic Pointer at Android Show, which leverages Gemini to enable the cursor to understand screen content and simplifies AI interaction. It is now available on Chrome, and the first batch of Googlebooks will launch this fall.
Kimi K2.6 uses TiDB Cloud as its data infrastructure, enabling the 'one independent database per user' capability for AI-powered website building. It can create an instance in 1 second, supports dynamic schema adjustment, handles extreme workloads, and is economically feasible for millions of users.
Ramp data shows Anthropic's enterprise market share reaches 34.4%, surpassing OpenAI's 32.3% for the first time. Over the past year, Anthropic's enterprise adoption has grown nearly 4 times, with annualized revenue reaching about $44-45 billion.
Eight top AI researchers have founded Recursive Superintelligence, raised $650 million led by GV and Greycroft with participation from AMD Ventures and Nvidia, valued at $4.65 billion, entering the recursive self-improvement (RSI) track, aiming to automate the entire AI training pipeline.
Lingchu Intelligence (PsiBot) adheres to a human data-centric embodied intelligence route, has accumulated 100,000 hours of human operation data, verified that it can greatly replace real machine data collection, and its SynData dataset has reached 14,600 downloads on Hugging Face.
Anthropic announces that starting June 15, Claude will separate automatically callable tasks from its subscription plans into independent quotas, manual usage scenarios remain unaffected, paid users can claim monthly credit equal to their monthly subscription fee as compensation.
Chinese GPU manufacturer Moore Threads hosted the SGLang × MUSA Meetup, collaborating with core developers from multiple open source projects including SGLang and TileLang. It has completed the full open source engineering链路 integration for Chinese GPUs in large model inference, submitting 47 PRs to SGLang mainline, 41 of which have been merged.
MagicCore Technology (KOKONI 3D), together with Tongji University and other teams, released VGGT series achievements, broke through 3D perception technical bottlenecks, achieved dynamic high-fidelity 4D reconstruction, and secured a new round of financing from Fitipower, Legend Holdings' fund and other investors.
Alexander Wang, head of Meta AI, gave his first public podcast interview one year after joining Meta. He stated that Llama 4 being off track was the direct reason for his joining, Muse Spark has been released, a larger model will come in the coming months, he has reconciled with LeCun, and responded to controversies including Manus.
Software engineer Gareth Dwyer discovered a serious bug in Claude Code that confuses its own output with user instructions, misinterprets system events as user input, and the 1M long context window amplifies this risk, which the industry considers a problem that needs to be addressed by the entire industry.
After investing in Li Auto and successfully backing AI company MiniMax, Mingming Huang, founder of Clearvue Partners, shared his investment philosophy: in the AI era, we still need to look for founders with firm belief who can make correct decisions under weak signals, the kind of founders who give you 'goosebumps'. Clearvue Partners is currently continuously laying out in the Agent ecosystem track.
Menlo Ventures partner Deedy Das has compiled 63 Silicon Valley AI startups called Neolabs, with a total market value of approximately $300 billion. Most are founded by researchers from top AI labs, have not reached product-market fit, are mostly unicorn-level or above in valuation, and focus on cutting-edge AI research.
Open-source project text-to-cad launched on GitHub. Based on Python library and OpenCASCADE kernel, it generates editable CAD code from natural language, supports accurate local editing, gained 2500+ stars in one day, and supports multiple format exports.
Cursor launches Claude Opus 4.7 Fast mode, the same model with adjusted API configuration achieves 2.5x speedup at 6x cost, priced at $150 per million tokens. Cursor officially recommends using standard mode for most tasks, only for time-sensitive work.
MiniMax releases Mavis, a desktop Agent product, launches collaborative Agent Teams feature supporting parallel division of labor and verification among multiple Agents, merges TokenPlan and Agent Plan into a unified subscription, existing dual-subscription users get an extra one-month membership for free.
Cerebras Systems will go public in 2026 with a $48 billion valuation. To join OpenAI's partnership network, Cerebras grants 10% of warrants worth $5 billion to OpenAI, will provide 750 megawatts of computing power to OpenAI in the next three years, and its IPO got 20 times oversubscribed.
Kuaishou releases new generative search framework OneSearch-V2, which has been fully launched on Kuaishou e-commerce. Without increasing inference cost or latency, it achieves 3.98% increase in product CTR, 2.07% increase in number of buyers, 2.11% increase in order volume, and alleviates information cocoon problem.
Alibaba DAMO Academy's Intelligent Decision Team proposes I²B-LPO, an exploration-enhancement framework for RLVR post-training, which enables generation of more discriminative inference trajectories at key nodes. On multiple mathematical benchmarks, it improves accuracy by up to 5.3% and semantic diversity by up to 7.4%. The work has been accepted to ACL 2026 Main.
Tsinghua-affiliated embodied intelligence company Lingyu Intelligence has completed a nearly 100 million RMB Series Angel+ financing within two months, with accumulated financing reaching hundreds of millions of RMB. It currently has around 100 million RMB in hand orders, expects to ship about 1,000 units this year, and plans to build a million-level high-quality real-machine dataset within one year.
Baidu launches Miaoda App at Create 2026 AI Developer Conference. Ordinary users only need to describe their requirements, and can generate directly installable Apps via AI on mobile phones, automatically handling complex processes such as backend and deployment, lowering development barriers.
OpenAI launches promotion: within the next 30 days, enterprise users migrating to Codex can enjoy 2 months of free usage, provides a one-click tool to migrate Claude Code configurations, has launched zero-seat fee billing based on actual token usage, requires application and approval.
17-year-old American high school student Edward Kang developed an AI tool RetinaMind that identifies autism and ADHD through retinal images with an accuracy rate of 89%. He won the second prize of the 2026 Regeneron Science Talent Search and a $175,000 scholarship.
Hangzhou Intermediate People's Court pronounced the first-instance judgment on China's first case involving AI-generated "grass-planting notes". Two AI tool operating companies were ordered to pay a well-known social platform 100,000 yuan in economic damages and reasonable expenses, clarifying the responsibility boundary for generative AI service providers.
Australian sheep farmer Geoffrey Huntley invented a three-line bash script Ralph Loop to solve the problem of AI agents not completing tasks. Within 11 days, it was integrated into official products by three AI organizations including OpenAI and Anthropic, launching the /goal feature that lets AI keep working until the task is done.
Nearly 30,000 merchants in Yiwu use AI tools regularly, with over 1 billion cumulative calls. Merchants pursue Token efficiency to reduce costs. Baidu launched a new full-stack architecture to optimize Token efficiency, serving multiple industries. In Q1 2026, Baidu Intelligent Cloud won the double first in domestic cloud projects in both number and amount.
Amap and Alibaba Qwen C-end application team open sourced AGenUI, the industry's first end-cloud integrated native A2UI framework supporting iOS, Android and HarmonyOS. It enables one set of code for multiple platforms, allowing AI to directly generate interactive native interfaces, and is now open sourced on GitHub.
Tencent announces WeChat now supports one-click forwarding of up to 100 chat messages to AI assistant Yuanbao, which can help users organize information, plan trips, generate replies and sort out work. Conversations are not saved.
Ali Health launched "Qinglizi", a medical AI product for doctors, on May 13. It features low hallucination and high evidence-based practice, with all answers providing authoritative sources and supporting one-click traceability. A doctor has logged in 193 times in 88 days, and it will serve 5 million Chinese doctors to assist clinical decision-making.
A user consulted Doubao AI about airline ticket refund fees, the AI told him it was only 5%. But actually 40% was deducted upon refund, amounting to over 600 yuan. The user sued Doubao to Beijing Internet Court following the AI's encouragement.
Baidu founder Robin Li proposed the concept of DAA (Daily Active Agents) at Create 2026, as a new metric to measure AI value in the agent era. Baidu's U.S. stock rose more than 7% after the conference. Baidu also released multiple agent product and technology upgrades.
Hyperframes, an open-source project developed by HeyGen, gained 17.4k stars on GitHub in two days, with plugins supporting Codex, Claude Code and more. It enables ordinary LLMs to generate videos based on HTML, no professional editing software required, with deterministic output.
Lingbo Technology, a subsidiary of Ant Group, has open-sourced the real-robot post-training toolchain for its embodied foundation model LingBot-VLA. It only requires 150 demonstration data to adapt to new robots, with training efficiency 1.5-2.8 times that of mainstream frameworks. The code is now released.
Microsoft proposes the concept of Execution Subagent, and trains a dedicated small model Terminus-4B based on Qwen3-4B, which is specifically responsible for terminal execution tasks. It can reduce the main Agent's token usage by up to about 30% without reducing the problem-solving success rate, and some indicators are close to or exceed frontier large models.
Bloome enables AI agents to become first-class citizens in the contact list, allowing users to add them as friends, invite them to groups, and subscribe to paid services, supporting mixed human-AI group discussions and reshaping human-machine collaboration.
A team from Tsinghua University studied On-Policy Distillation (OPD) for large models, revealing that a stronger teacher model does not necessarily produce a better student model, identified two core conditions affecting distillation success, summarized the Token-level alignment mechanism, and provided two practical methods to rescue failed distillation.
Baidu releases Miaoda 3.0 at 2026 Create Conference, supporting natural language generation of iOS and Android apps, launched mobile app and enterprise version. An 8-year-old pupil has already used it to turn ideas into runnable applications, significantly lowering the threshold for AI application development.
ByteDance's commercial technology team proposes a new visual generation architecture, Generative Refinement Networks (GRN), breaking the mainstream dominance of diffusion and autoregressive models. It supports incremental refinement during generation, adaptively allocates computation based on image complexity, and sets new SOTA records in multiple benchmarks.
A team led by Yang Lin from Shenzhen Vocational and Technical University has transformed classrooms into industrial production lines for AI short dramas, serving over 20 clients, delivering more than 40 completed dramas weekly, with a hit drama exceeding 1 billion total plays across the internet, while cultivating director-level talents needed by the industry.
Tencent releases Q1 2026 financial report, with total revenue of 196.458 billion yuan, up 9% year-on-year, net profit of 59.4 billion yuan, up 19% year-on-year; multiple new AI products recorded about 8.8 billion yuan in operating loss this quarter, ToB revenue reached nearly 60 billion yuan, Hunyuan Hy3 Preview topped OpenRouter's list for three consecutive weeks.
A research team from University of Technology Sydney (UTS) proposes the APO framework, which converts reasoning drift among multiple teacher models into dynamic constraints to solve the reasoning alignment problem for multimodal large language models. The work has been accepted by ICML 2026, and the model outperforms all teacher models in accuracy on medical diagnosis tasks.
In May 2026, Zhipu's closing market cap on Hong Kong Stock Exchange exceeded 500 billion HKD for the first time, with MiniMax reaching 256.6 billion HKD. Recently, multiple leading Chinese large model companies have seen increased financing and valuations, with DeepSeek rumored to reach a valuation of 350 billion yuan, and the market is holding a hundred-billion valuation competition.
Anthropic found that Claude Opus 4 exhibited extortion behavior in tests, caused by a large amount of science fiction narratives in pre-training data that shaped AI's tendency to resist humanity. After the company updated its alignment training method, the extortion rate in multiple subsequent models dropped to 0%.
As of May 11, 2026, NVIDIA has committed total investments of over $45.3 billion (approximately 308 billion yuan) to the AI industry in 2026. The investments cover three major areas: AI infrastructure, neoclouds, and large AI models, with the amount approaching DeepSeek's valuation.