The takeaway: China spent the last 24 hours building the institutions of AI governance — 29 countries signed the World AI Cooperation Organization into existence in Shanghai on Thursday, and Xi Jinping opened the World AI Conference with a keynote this morning. Meanwhile Moonshot AI launched Kimi K3, a 2.8-trillion-parameter model that arrived with "open-source" edited out of its launch copy mid-launch — and a dated promise to publish the weights by 27 July. The frontier and its politics, in the same news cycle.

A new AI organisation, headquartered in Shanghai

Twenty-nine countries signed the agreement establishing the World AI Cooperation Organization (WAICO) in Shanghai on Thursday, according to Xinhua. The body is described as "an independent intergovernmental international organization headquartered in Shanghai" — a permanent, treaty-based institution, not a talking shop. Chinese Foreign Minister Wang Yi signed on behalf of China, and UN Secretary-General António Guterres was among those present at the ceremony.

Xinhua names Kazakhstan, Laos, Pakistan, Russia and Indonesia among the founding members; the full list of 29 signatories has not yet been published, and no Western signatures have been reported. That is the honest shape of it: a China-anchored governance bloc, formalised the day before the World AI Conference opened.

Then, this morning, Xi Jinping delivered the keynote at the opening ceremony of the 2026 World AI Conference and the High-Level Meeting on Global AI Governance in Shanghai — his first in-person WAIC address. As we publish, Xinhua has confirmed the speech happened but not released the text. We will report on the transcript when it is up, not on second-hand summaries — expect a follow-up today.

For the run-up to this moment, see yesterday's Daily Update.

Kimi K3: 2.8 trillion parameters and a promissory "open"

Moonshot AI launched Kimi K3 on Thursday afternoon — a 2.8-trillion-parameter Mixture-of-Experts model (16 of 896 experts active, built on the company's Kimi Delta Attention and Attention Residuals) with native vision and a 1-million-token context window. The company calls it "the world's first open 3T-class model." Its Hacker News thread passed 1,200 points overnight.

But the word "open" did some travelling on launch day. Readers with the original launch page still open in their browsers preserved its first claim: "Kimi K3 is the first open-source model to reach the 2.8-trillion-parameter scale." That paragraph was then removed — one commenter noted mid-afternoon that "any mention of 'open' has been struck from the literature for this model (it was present an hour ago)." By evening, the language had settled where it stands now: K3 is "the first open model to reach 2.8 trillion parameters," and "the full model weights will be released by July 27, 2026." No licence has been named. Nothing has appeared on Hugging Face.

Until 27 July, then, the open crown is promissory. Artificial Analysis — whose independent scorecard is the one to watch — currently classes K3 as a proprietary model ("the model weights are not publicly available") and scores it 57 on its Intelligence Index, #4 of 189 models tracked, with only three proprietary frontier models above it. Moonshot's own launch copy is unusually candid on that point: K3 "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol."

Early hands-on reaction runs hot and cold in the same thread: one developer's early hands-on update reported "the subscription limits are pretty brutal" but "the model itself is amazing. I think I might put this above Opus 4.8" — while a sceptic countered that "the benchmarks aren't going to tell the whole story," suspecting benchmark leakage into training data. API pricing is $0.30 per million tokens for cache-hit input, $3.00 for cache-miss input and $15.00 for output.

We are working on a full K3 piece today — the edit saga, the independent numbers and what "open" means when the weights are still ten days away. Our field guide to what 'open' actually means in AI is the primer.

In brief

• No Gemini 3.5 Pro. Today was the rumoured launch date; as of this morning there is nothing official from Google — the Gemini API changelog's latest entry remains 6 July. Rumour until Google posts, and Google has not posted.
• DeepSeek's deadline week. DeepSeek discontinues its legacy deepseek-chat and deepseek-reasoner model names on Friday 24 July — a week today — completing the V4 transition that began in April. Anything still pointing at the old names has a week to move.
• Nobel laureates on AI and the bomb. At the Global Nobel Laureates Assembly at the Vatican, signatories including Juan Manuel Santos and Denis Mukwege issued "Humanity at the Threshold," a declaration on AI and nuclear weapons published on Thursday: "An automated system should never make the final decision to launch a nuclear weapon." It also calls for a treaty against putting AI into nuclear command-and-control, and for a "coordinated slowdown of frontier AI development" — a striking counterpoint to WAICO's development-first vision, landing the same day. More on this later today.
• Fable promo ends Sunday night. Anthropic's Claude Fable 5 promotional access ends Sunday 19 July at 11:59pm Pacific time — 07:59 on Monday morning, UK time. Our guide to the cliff has the details and the decision framework.