AI News Weekly: Regulations Tighten, Models Surge, Robots Enter the Factory
The first week of July 2026 delivered a firehose of AI news — from policy crackdowns on companion bots to trillion-parameter model launches, humanoid robots entering real factory lines, and a wave of funding that shows no sign of slowing.
Here’s what happened and why it matters.
🚨 Regulation Bites: China Shuts Down AI Companion Features
Alibaba’s Qianwen and ByteDance’s Doubao will remove their “agent” features on July 15, following notices to users that agent configurations and historical chat logs will become inaccessible. The affected features — AI chatbots built on knowledge bases for roleplay, English tutoring, and emotional companionship — fall directly under the “Interim Measures for the Management of AI Anthropomorphic Interactive Services” that takes effect on July 15.
This marks a real turning point. China’s AI companion industry, which saw explosive growth in 2025-2026, is now navigating its first major regulatory contraction. ByteDance is directing Doubao users to migrate to a standalone app called “Catbox” to continue using companion services, suggesting companies are segmenting their products to keep companion features alive under separate regulatory frameworks.
🧠 Model Landscape: Trillion-Parameter Era Accelerates
Meituan LongCat-2.0: 1.6 Trillion Parameters on Domestic Silicon
Meituan released LongCat-2.0, a Mixture-of-Experts model with 1.6 trillion total parameters and ~48 billion activated per token. It natively supports a 1-million-token context window and, crucially, is the first trillion-parameter model to run its entire training-to-inference pipeline on domestic Chinese hardware. Meituan has fully cut off external large model usage, going all-in on its own stack.
Anthropic Claude Sonnet 5: Production-Ready Agent
Anthropic launched Claude Sonnet 5, a model capable of running long-duration autonomous tasks — browsing, using a terminal, executing multi-step workflows. Performance is close to Opus 4.8 but at a significantly lower price point. Separately, Anthropic lifted US export controls on Fable 5 and Mythos 5, restoring global access. The company also launched Claude Science, a research workbench integrating 60+ specialized skills in genomics and proteomics — benchmarked at the level of a second-year graduate student.
Kimi K3 Coming in July: 2.5 Trillion Parameters
Moonshot AI’s Kimi K3, expected this month, will pack 2.5 trillion parameters with 1M-token context and deep multimodal integration. Kimi’s annualized recurring revenue has crossed $300M, with API revenue accounting for over 70%.
Meta Enters Cloud: “Meta Compute” Infrastructure
Meta is building “Meta Compute”, a cloud infrastructure business that will directly sell AI compute and model access to external customers — putting it in competition with AWS, Microsoft, and Google Cloud. The move comes as Meta raises its 2026 AI CapEx to 145B.
Google’s Tiny Fast Models
Google released Nano Banana 2 Lite, an image model that generates 1K-resolution images in 4 seconds at $0.034 per image, and Gemini Omni Flash, a video model supporting conversational video editing. Both are integrated into Gemini App, YouTube, and other Google products for free.
OpenAI GeneBench-Pro
OpenAI released GeneBench-Pro, a biology benchmark covering 10 core domains and 129 tasks. Their new flagship model, Sol, outperformed Claude Mythos 5 on the Terminal-Bench 2.1 benchmark.
🤖 Agents Go Mainstream
Alibaba is consolidating three enterprise agent products — QoderWork (desktop tool), DingTalk’s “Wukong”, and Alibaba Cloud’s “MuleRun” — under 33-year-old Chen Yusen to build an “OS-level” desktop agent.
Alipay’s “A-Bao” AI life assistant opened public beta with 72 skills covering ride-hailing, food delivery, government services, and more, accessible via text or voice.
WeChat launched an A2A assistant capability with phone manufacturers — one command to send a WeChat message. Public accounts on WeChat can now deploy AI avatars for 24/7 customer service.
Feishu (Lark) introduced a multi-dimensional table agent, upgrading AI from personal assistant to team-level “AI colleague” embedded in group chats and spreadsheets, capable of natural-language multi-source queries and proactive business monitoring.
xAI’s Voice Agent Builder is a no-code platform that configures a production-grade Grok Voice agent in ~2 minutes, priced at $0.05/minute.
Cursor disclosed two critical vulnerabilities (“DuneSlide”) that bypass its sandbox for system-level control. Meanwhile, Cursor launched an iOS app for remote agent programming.
🦾 Physical AI: Robots Leave the Lab
Tesla Optimus Enters the Factory
Elon Musk confirmed that Tesla Optimus humanoid robots have begun work on the Fremont factory production line. Initial production is slow — official mass production is expected in late July or August, with large-scale rollout in 2027.
NVIDIA Open-Sources ASPIRE Skill Library
NVIDIA released ASPIRE, an open-source robot skill library — think of it as a “coding agent for robots.” It accumulates failure and recovery experiences into reusable skills. Bimanual object handoff success rates jumped from 20% to 92% using ASPIRE.
Ubtech and Japan’s Robot Push
UBTECH’s U1 hyper-realistic humanoid robot series secured over 10,000 unit orders, powered by an embedded emotional companionship AI model. Japan announced a plan to deploy 10 million robots by 2040 to address its aging population crisis, along with developing a sovereign large language model.
💻 Hardware & Infrastructure
Nvidia optimized DeepSeek V4 inference on Blackwell, cutting per-token cost by up to 5x and boosting throughput by 20x.
Cambricon briefly hit a $140B market cap, becoming the first STAR Market stock to surpass 1 trillion RMB before issuing a risk warning and pulling back.
SiliconFlow filed for a Hong Kong IPO — the “AI Token Factory” has 10M+ registered users and peak daily token throughput exceeding 1 trillion.
Nvidia became the #1 data center ethernet switch vendor by revenue in Q1 2026, expanding from GPU compute into networking infrastructure.
💰 Money Keeps Flowing
Kling AI is closing a **18B valuation, with Tencent participating. IPO in Hong Kong targeted within 12 months. Q1 2026 revenue exceeded $650M.
SoftBank invested an additional **10B tranche by October 1, 2026.
Etched (AI inference chip startup) hit **5B, with $1B in annual sales. First rack-scale products ship this summer.
Blackstone plans to invest $30B in Japanese AI data centers over 3–5 years, adding 1GW+ of compute capacity.
📅 WAIC 2026: What to Watch
The 2026 World AI Conference (WAIC) runs July 17–20 in Shanghai under the theme “AI Partners, Co-Creating the Future.” This year features the inaugural “WAIC Academic” conference chaired by Turing laureate Andrew Yao, with over 100,000 sqm of exhibition space, ~140 forums, and 300+ AI products making their global debut.
References
- Caixin: Alibaba/ByteDance AI companion feature shutdown (internal reporting)
- Meituan LongCat-2.0 announcement
- Anthropic blog: Claude Sonnet 5 launch (anthropic.com)
- Moonshot AI: Kimi K3 preview (kimi.moonshot.cn)
- Meta: Meta Compute infrastructure plan (meta.com)
- Google: Nano Banana 2 Lite / Gemini Omni Flash release notes
- OpenAI: GeneBench-Pro announcement (openai.com)
- Alipay: A-Bao public beta (alipay.com)
- WeChat: A2A assistant capability launch
- Feishu: Multi-dimensional table agent blog (feishu.cn)
- xAI: Voice Agent Builder (x.ai)
- Cursor: DuneSlide vulnerability disclosure (cursor.com)
- Tesla: Optimus factory update (tesla.com)
- NVIDIA: ASPIRE open-source release (developer.nvidia.com)
- UBTECH: U1 order announcement (ubtech.com)
- Cambricon: Market cap milestone (cambricon.com)
- SiliconFlow: HK IPO filing (siliconflow.cn)
- Nvidia: Ethernet switch market share report
- Kling AI: $3B funding round report
- SoftBank: OpenAI investment (softbank.jp)
- Etched: Funding announcement (etched.com)
- Blackstone: Japan data center investment (blackstone.com)
- WAIC 2026: Official program (waic.sh.cn)