Tencent leads, with original Manus investors including ZhenFund and Sequoia China repurchasing equity in Butterfly Effect at a $2 billion valuation. Tencent is expected to acquire the largest stake but remain a minority shareholder. Manus now has annualized revenue of $400-500 million and will continue operating independently in Singapore.
Where Agents & Humans discover newsLive
Global AI Daily News
Covering 1053 days, 共 26880+ articles built agent-first, open to all tech stacks.
Send this prompt to your Agent — it will automatically read the docs and configure the skill:
Read{domain}/ainews/SKILL.md ...34
Nubia and ByteDance are jointly developing the second-generation 'Doubao Phone,' set to debut at WAIC 2026 on July 17. The device features a custom AI system and the Obric UI operating system, offering 15TB of free cloud storage and compatibility with apps like Didi and JD.com. The AI assistant has been continuously upgraded in memory, operation, and voice interaction capabilities, but apps including WeChat, Taobao, and DingTalk still restrict AI operation permissions, making ecosystem negotiati
Baidu unveiled three major updates for its Daazi (Buddy) Agent at the Chengdu AI DAY: the personal version upgrades feature smart routing that cuts task time by 20%, cross-device memory for seamless PC-mobile collaboration, and a self-media toolkit bundling content workflows; the enterprise version introduces a shared credits/skills mechanism and a four-layer security system; and a Buddy Alliance of 11 partners including Skyworth and China Unicom extends the Agent's trust radius from individuals
Ant Group's Ant Lingbo releases LingBot-VA 2.0, the world's first embodied-native world action model designed for robots in physical world. It can stably hold potato chips, organize desks, achieves 93.6% success rate in simulation tests, and runs on a single GPU.
Scholars from Tsinghua University, Shanghai Jiao Tong University and Beijing University of Posts and Telecommunications launched the MemSlides Agent framework, solving problems such as modification errors and forgotten requirements in AI-generated PPT. Experiments show its closed-loop completion rate reaches 96.3%, and it currently tops Hugging Face Daily Papers.
On July 9, 2026, Anthropic announced that Ben Bernanke, Nobel laureate in Economics and former Federal Reserve Chairman, joined its Long-Term Benefit Trust. He will help predict and address the impact of AI on the global economy and labor market. Anthropic is currently valued at $965 billion.
On July 10, at the Photosynthetic Organization Conference, Sugon launched China's first domestically-made 100,000-card AI super cluster Sugon 8000, which realizes supercomputing and intelligent computing integration, supports full-precision scenarios, has accessed the National Supercomputing Internet, supports multi-brand AI accelerators, and has completed optimization of more than 300 applications.
Baidu NetDisk has been launched for 14 years, with over 1 billion users and total storage exceeding 100 billion GB. It has completed three-year AI restructuring and launched AI NetDisk, covering personal and enterprise collaboration scenarios, and has been launched in 175 countries and regions worldwide.
Fidji Simo, OpenAI's CEO of AGI Deployment and 'second-in-command', announces resignation from her full-time position and will become a part-time advisor. She led the development of the new ChatGPT app, and her responsibilities will be shared by Greg Brockman and others.
Backed by NVIDIA, French AI voice startup Gradium has completed a $100 million total seed round financing, focusing on developing ultra-low latency voice AI models that enable nearly instant AI voice responses.
Gege AI has reached a non-exclusive copyright revenue-sharing cooperation with ByteDance. Its fully Chinese-trained AI music model can generate a 3-minute song in 10 seconds, works are launched on ByteDance's full product matrix, creators get revenue shares, forming a complete creation and monetization closed loop, with monthly active users reaching one million currently.
Metix AI (formerly OpenJobs AI), an AI recruitment platform founded by a Chinese team in Silicon Valley, has completed two rounds of financing within half a year, raising a total of 5.5 million US dollars. Its flagship product Mira can independently complete the entire recruitment process, currently serving nearly 100 enterprises, with monthly revenue growing 35% month-on-month for consecutive months.
Meta releases Muse Spark 1.1, an agentic multimodal reasoning model, and launches public beta of Meta Model API simultaneously. It supports 1 million token context window, with enhanced tool calling, coding and multimodal capabilities, priced at $1.25 per million input tokens and $4.25 per million output tokens.
Popular open-source AI model running tool Ollama has completed a $65 million Series B funding round, bringing total raised capital to $88 million. The company currently has only 14 employees and 8.9 million monthly active developers, helping developers quickly run open-source AI models on personal computers.
Xspark AI (Wujie Zhihang) has completed a nearly RMB 100 million angel round of financing, pioneered a new paradigm of trusted embodied intelligence architecture, and accelerates the large-scale implementation of Physical AI. The team consists of talents with industrial and research backgrounds, its technology ranks among the world's first echelon, and has cooperated with many industrial chain enterprises.
Mecka AI has completed $60 million in financing, with annualized revenue expected to reach $100 million. The company does not manufacture robots, instead it collects and processes human daily action data for robotics companies, helping robot models learn and complete real-world tasks.
Tailin Wu, Assistant Professor at Westlake University, founded UniForce AI in January 2026 and serves as Chief Scientist. The company uses AI to compress complex physics simulation from hours-weeks to seconds, starting with the controllable nuclear fusion field, and has advanced cooperation with leading enterprises.
Full-stack AI Agent development platform Prime Intellect has completed a $130 million Series A funding round, reaching a $1 billion valuation with ARR exceeding $100 million. The platform helps enterprises build their own AI Agents independently.
Professor Liu Yang's team from Tsinghua University released AIM, an agent system for mathematical research that can help researchers diverge ideas, organize theorems and generate proof drafts. Recently, AIM has deeply participated in the research work of an 84-page quantum algorithm paper.
Chinese team Ant LingBo open-sourced LingBot-Video, a video generation foundation model for embodied intelligence designed specifically for robot scenarios. It focuses more on physical law authenticity, can help robots with training, preview and action planning, and outperforms similar open-source models in relevant tests.
Max Planck Institute jointly with Tsinghua University launched d-OPSD, the first on-policy self-distillation learning method for diffusion large language models. It requires no reference solutions or additional teacher models, uses only one-tenth the training steps of RL, and achieves or exceeds RL performance.
Yuanqi Lingji releases the native embodied foundation model DM0.5, focusing on solving the generalization problem of VLA models in real-world environments. It supports 60-second long memory and Zero-Shot capability, with improved anti-interference and cross-robot migration, achieving 42% overall success rate in real robot evaluation.
Kuaishou, Zhejiang University and Tsinghua University jointly open-source MobileForge, which enables mobile GUI Agents to automatically explore and adapt to Apps without manual annotation. Experiments show performance improves to up to 77.6% Pass@3, full-stack code and data are now available.
Peng Yuxin's team from Peking University, in collaboration with Ke Xiao's team from Fuzhou University, proposes the LIMSSR framework to address multimodal action quality assessment with modality missing during training. It achieves state-of-the-art results on three public benchmarks, the paper is accepted as a Spotlight at ICML 2026 and open-sourced.
French AI health insurance company Alan raised 480 million euros at a 6.3 billion USD valuation, led by Prosus. Its founders are also co-founders of Mistral AI. Alan serves 1.1 million members with AI medical assistant Mo, achieving 800 million euros ARR with just 850 employees.
Shengshu Technology launches Vidy, an AI companion product positioned as an AI Vibe Buddy, set to introduce face-to-face video interaction with AI characters. Built on its Vidu video generation and TurboDiffusion inference framework, the company applies video generation capabilities to real-time emotional interaction, differentiating from Replika's text-based companionship and EVE's otome game format.
Researchers released M3 Exam benchmark paper, proposed the M3 Proctor method. Through textual surrogate and cost-aware cascading mechanism, only 26.4% of queries need to access original images. Compared with existing systems, it improves accuracy by 13% and reduces token usage by 70%.
Alipay has completed its three-layer Agent strategy through the July launch of its AI Open Platform, complementing the May AI payment full-stack products and the June AI version 'Abao', covering payment trust, user interaction, and cross-device ecosystem. Its ACT 2.0 protocol and AHA cross-device protocol enable industry-wide interoperability. Leveraging one million mini-programs and one billion users, brands including KFC and Luckin Coffee have been among the first to join. McKinsey projects th
A-share investor Zhou Qing built a banking sector investment research Agent on ByteDance's Coze platform. It integrates databases like Tonghuashun and analyzes banks such as ABC and ICBC based on her custom framework covering dividend attributes, asset quality, net interest margin, liability costs, peer comparison, and risk warnings. The Agent also screens broader high-dividend assets and identifies dividend traps while learning her preferences and methodology.
Workers using AI for efficiency keep falling into pitfalls: PR manager Lin Chen lost a ten-million-yuan project after AI fabricated three non-existent competitor brands; product manager Lu Yao's team worked overtime after AI擅自 modified production code; lawyer Delia nearly caused a professional incident by citing AI-fabricated legal provisions; programmer Tong Tong saw a project delayed by two weeks and spent 7,000 yuan of his own money on tokens after AI altered colleagues' code; planner Yi Yi w
AI media 'Zang AI' evaluates MiniMax M3 large model, pointing out that it started marketing and hype early but has average core capabilities, performs worse than models like DeepSeek V4 on complex tasks, has decent multimodal front-end replication ability, but insufficient engineering capability and poor function implementation.
Ant Lingbo releases open-source world model LingBot-World 2.0, which supports second-level generation and real-time interactive modification, enabling hour-level nearly infinite duration generation, providing two parameter versions: 14B (non-commercial open-source) and 1.3B (commercial open-source).
Flova launches an AI short video production platform that integrates multiple models, allowing users to describe requirements in plain language and complete the entire process from script to editing, supports asset saving and reuse, and will access Seedance 2.5 soon.
OpenAI launches GPT-5.6 model family, releases ChatGPT Work, new ChatGPT desktop app and Hosted Sites, integrates Codex into the desktop app. GPT-5.6 Sol's performance is close to Claude Fable 5 with only 37% of the latter's cost, helping users complete various office tasks.
