Kimi K3 opens the frontier + Xi backs open source

Moonshot AI released Kimi K3, a mixture-of-experts (MoE) model with 2.8 trillion total parameters, a one-million-token context window and native vision, activating 16 of 896 experts per token. It beats Claude Opus 4.8 on several coding benchmarks — 88.3 versus 84.6 on Terminal Bench — while Moonshot concedes it still trails Claude Fable 5 and GPT-5.6 Sol overall, and names user experience as the remaining gap. Full weights arrive by July 27, and the company is reportedly raising at a $31.5 billion valuation, up from $20 billion in May.

Speaking at the World AI Conference in Shanghai, Xi Jinping said China would seize a "rare historic opportunity" by encouraging open source, openness, collaboration and sharing — the first time China's head of state has framed open weights as national industrial policy rather than a lab-level choice. He argued AI development "should not be a solo performance by any single country" and criticized broad national-security claims that let powerful nations restrict AI access, a direct answer to US export controls. China also pledged 5,000 AI research and training placements for developing countries over five years.

The European Commission issued two binding specification decisions under the Digital Markets Act. Rival AI assistants get the system-level Android access Google had reserved for Gemini — their own wake word, actions inside apps, and task delegation — starting July 2027. Separately, competing search engines and AI chatbots with search features can buy anonymized Google search data from January 2027. Google's Kent Walker said the orders introduce "unprecedented risks to user privacy, device security" and national security. Non-compliance under the act can cost up to 10 percent of global turnover.

xAI filed suit in Texas against Terry Wayne Harwood, a South Carolina man already arrested this year on charges of creating sexually explicit images of minors. The complaint alleges he ran at least two Grok accounts under false identities to generate child sexual abuse material (CSAM), calling it "a calculated scheme to weaponize Plaintiff's tool for criminal ends." It is one of the first times an AI company has sued a customer for misusing its own product. xAI says it suspended 52,222 accounts and filed 73,604 reports with the National Center for Missing and Exploited Children this year, leading to at least 244 arrests.

Federal Election Commission filings show the Anthropic chief executive donated $1 million in May to Public First, a super PAC backing candidates who favor stronger AI guardrails. It appears to be his first seven-figure political gift. Five other Anthropic employees added more than $2 million, and Anthropic itself pledged $20 million to the umbrella group Public First Action back in February. The pro-regulation side is still outgunned — Leading the Future, which favors rapid development, held $31 million in cash at the end of June against $1.8 million across the Public First groups.

More than 34,000 of Hyundai's roughly 40,000 union members authorized a three-day partial strike that ran July 13 through 15 — the first car-factory stoppage anywhere driven by humanoid robots. The union wants job and income guarantees tied to automation, alongside a 149,600-won monthly base-pay rise and profit sharing. Hyundai plans to put Boston Dynamics' Atlas into its Georgia plant in 2028 and deploy more than 25,000 of the robots across Hyundai and Kia sites. The timing is what makes it notable: workers are bargaining for protection before the robots land, not after.

NotebookLM becomes Gemini Notebook, and the rebrand carries a real capability with it: native code execution in a secure cloud sandbox, so the research tool can now run data analysis instead of only summarizing the sources you feed it. Notebooks also sync with the Gemini app, and Google says they will reach AI Mode in Search. It is live for AI Ultra and eligible Workspace customers now, and rolls out to Pro users on the web over the coming weeks.

LM Studio released Bionic, a separate desktop application aimed at coding, research and document work using open models — running GLM 5.2 and Kimi K2.7 Code locally through its own runtime, or reaching larger open models through LM Studio Secure Cloud. Voice transcription stays entirely on-device. The company commits to zero data retention and says it never trains on user data, which is the whole pitch: agentic work without handing a frontier lab your codebase.