🔴 China Has Forced Its Way into the Frontier AI Club

Masterful timing. Moonshot launched Kimi K3 on July 16, on the eve of the World AI Conference in Shanghai, where Xi Jinping delivered the opening speech. An opportunity for the Chinese leader to announce that he would bring the benefits of artificial intelligence to the Global South. A speech at the opposite pole from that of the United States and from the United Nations' at the AI for Good summit.

The weights are the model itself: the file that contains everything the machine has learned. Once published, anyone can download them. But running them is another matter: a model of this size demands considerable infrastructure. Only a lab or a company equipped with clusters will be able to run it on its own machines, modify it, specialize it, without asking Moonshot for anything, and without anyone being able to cut it off remotely.

That is the main drawback of an open-weight model.

There are plenty of fantasies about what a world that sets frontier intelligences free will look like, but also a new reality, documented by Hugging Face.

Cybersecurity, pilot fish of the super-frontier models

We have entered a new era. Cybersecurity has completely changed in the age of super-frontier models like Mythos. But no one knows what happens if open source models reach that level. Or almost no one.

In a study published on July 17, the UK's AI Security Institute estimates that Chinese open models now sit at the level of the best closed models from four to seven months ago. On targeted cyber tasks, GLM-5.2 matches Opus 4.6, released four months before it. On end-to-end autonomous attacks, it reaches the level of Opus 4.5, at under seven months. The gap, which was six to ten months in 2025, has therefore narrowed.

Hugging Face, the Franco-American platform that hosts open source models from around the world, was attacked in mid-July. A booby-trapped dataset (a deliberately corrupted set of data) made it possible to execute code on one of its servers, then to harvest credentials and spread across several clusters, over a single weekend.

Hugging Face's response illustrates this new era. To reconstruct the attack, more than 17,000 events had to be analyzed, impossible by hand given the speed. The team first tried to use the American frontier models through their commercial APIs. It did not work: their guardrails blocked the analysis, unable to tell an analyst's actions from an attacker's, because the data under examination contained exploit code.

The solution was to switch to GLM-5.2, the Chinese open-weight model, running it on their own infrastructure. The reconstruction took a few hours, at the attacker's pace.

Added benefit: no sensitive data from the investigation left their environment. This will be the standard model going forward.

Hugging Face is not just anyone. Running GLM-5.2 locally demands serious resources. The model contains 744 billion parameters, each stored on 2 bytes, roughly 1.5 terabytes to load entirely into memory. That takes a cluster of several GPUs, several hundred thousand euros' worth of machines.

Two lessons can be drawn from this incident.

• If your cybersecurity consultants do not have access to clusters running Chinese super-frontier models, they may no longer be able to help you in the event of an AI-driven attack.
• If you do not have open source models running directly, then you have to go through a provider's APIs, whose access, as we saw with Mythos, is now regulated by a state. Not to mention the confidentiality problem, because you have to send them your source code!

A headache for the entire French cyber sector.

We will come back a little further on, in the section for paying members, to what all this changes in American, Chinese and European strategies. We have reached this Kimi K3 moment 6 months ahead of schedule!

📚 To get up to speed, some of our newsletters have covered Chinese AI

• DeepSeek: An Artificial Intelligence to Arm the Rest of the World? (January 27, 2025): DeepSeek R1 as the Kalashnikov of hybrid and cyber warfare, a model immune to embargoes.
• Sovereignty, Cybersecurity, Cognitive Autonomy (June 1, 2026): Mistral's real place against the Chinese models, and the idea of a Westinghouse moment with Chinese AI.
• Embargo on Intelligence (June 15, 2026): the American frontier models put under a bell jar, and open routing reaching 99% of Fable 5's level for half the price.
• The Latent Space of the Middle (June 22, 2026): the big report on Chinese open models, Kimi K2.6, DeepSeek V4, GLM-5.2 and MiniMax M3.
• Will the Token Economy Hold Until the IPOs? (June 28, 2026): 60% of companies are switching to Chinese open models to cut their bill.

Become a paying member before prices go up

This newsletter is twenty minutes of in-depth thinking per week. Ideas you apply in your daily life. An intellectual exercise we all need to do in this period of rapid change. In two years, our members have become the de facto go-to people on questions of sovereignty, AI and the geopolitics of technology.

As a member, you receive:

• the full versions of the newsletters and their bonuses
• access to the site's complete archives
• all my strategic monitoring via a private channel
• priority answers to your emails

The low price is a choice, to allow students, teachers and small SMEs to subscribe. Many of you tell us that some notes alone are worth far more than the price of the subscription.

K3, who are you?