When China Gets a Mythos Model

The drop

Kimi K4 drops just as every model, Chinese or American, dropped before Mythos: a hype tweet. Only a few hours later, just as with its previous releases, Moonshot AI makes Kimi K4 open-source — weights available freely online. Well over a trillion parameters. No launch event, no licensing queue. A link.

Two discoveries, simultaneously

First, cyber professionals tinker around and discover Kimi K4 can find zero-day vulnerabilities the same way Mythos could. They scramble to inform labs and policymakers, both in the US and in China. Second, enterprises that just want good, cheap models realize Kimi K4 is a saving grace. Uber Eats, Siemens, and Shopify had already begun wiring Chinese models into their workflows — but now there is no trade-off between cheaper tokens and amazing performance. On the graph of enterprise usage of Chinese models, the slope takes a dramatic uptick.

Beijing’s dilemma

A model that democratizes cyber offense — something the Chinese government believed it had an edge in against everyone sans the NSA — is proliferating worldwide. Cyberattacks look poised to inflict disproportionate financial damage on China’s own leading companies, since America and friends have already been hardened by months of Mythos access.

But Beijing also sees the upside. Moonshot AI’s market cap approaches $500bn, instantly making it one of China’s largest companies. Kimi K4 is China’s chance to beat America in leading-model adoption, crater the western labs’ businesses, and become the go-to technology provider for global enterprises and governments. Besides, struggle is part of the process: the Party accepted sacrifices during COVID reopening, bank failures, and the tech crackdowns. This round of cyberattacks could just be the culling before China emerges stronger than ever.

Washington’s dilemma

The Trump administration is also torn. USG rails against China for distilling and then releasing such a model, declares that government contractors can’t build with Chinese open models — but doesn’t outlaw the models themselves; banning a file that can just be torrented would be silly. When Beijing sends no signals of a crackdown, Washington decides to let open source rip as well, screaming at its leading labs to drop something open. If there’s going to be a powerful open-source model out making waves, it should probably be American.

Inside Moonshot AI

Moonshot AI workers, already in a sleepless daze from the eighteen-hour days up to launch, don’t know what to do. The launch-party celebrations have subsided into stress. They are talking with the relevant authorities — but because they open-sourced the model, as they did with K3, there is no putting the genie back in the bottle. A model of well over a trillion parameters can’t be weaponized by anyone with too much free time and a Mac Mini. But there are plenty of actors — terrorists, companies with the wrong interests, criminal syndicates — who could muster the racks to turn it against China. All Moonshot AI can do is hold its breath and wait for the fallout.

The new ceiling

The Party decides Kimi K4 wasn’t dangerous enough to strangle — but the model did serve as a wake-up call. Kimi K4 is now the ceiling for unsupervised drops, and every model after gets scrutinized closely before release. Without much fuss, China implements a stricter licensing regime: one that checks not just for politically unacceptable output, but for cyber capabilities.

The hint

After seeing Mythos, Beijing realizes emergent capabilities could inflict real damage in the wrong hands. The best mitigation looks like a Project Glasswing-type undertaking. Ding Xuexiang gives a speech implying that breakout domestic AI models should be shared with important government ministries and corporate giants like Baidu, Alibaba, and Tencent before deployment.

The postponement

Yang Zhilin (杨植麟) and team had already deeply internalized the misuse risks of Mythos-level models. After Ding’s speech, Moonshot AI postpones the Kimi K4 launch party to evaluate the model internally — at which point its researchers realize what their creation is capable of. Determined not to be disappeared, Moonshot’s leadership pitches Project IceFeather (冰羽项目) to the Cyberspace Administration of China.

The announcement

The CAC, of course, takes the lead on announcing the initiative — with Moonshot AI credited as the project’s “core private-sector partner.” Government ministries and key SOEs like State Grid get access first, using their exclusive window to check their own systems for vulnerabilities, while the MSS trawls for exploits that can be used offensively overseas.

Endless evaluation

The Party’s involvement means the state is now co-owner of the model’s risks. For bureaucrats instructed to “evaluate first, deploy later,” the dominant strategy turns out to be endless evaluation. Letting anyone with enough compute punch holes in China’s critical cyber architecture seems far scarier to the system than marginally more productivity growth and global discourse power. At first, nobody quite decides the model should be released — nobody wants to be responsible for what happens post-release.

Despite holding the best Chinese model, Moonshot AI’s stock doesn’t moonshot: feet-dragging spooks investors who worry the government won’t let the lab monetize beyond state customers, much less overseas. While the MSS has the time of its life hacking everyone who didn’t get Mythos access, domestic coders still mostly use transfer stations to reach frontier capabilities.

The circle widens

After two months, regulators expand the trusted circle to provincial governments and trusted local tech leaders like Alibaba, Xiaomi, and Huawei. It takes another few months for Chinese firms broadly to get access — by which point other domestic labs have developed roughly similar models.

The trigger

With American models continuing to advance and the U.S.–China gap widening, the trigger gets pulled. Safeguards are put on the model — strong, perhaps to the point of excess — and regulators hope domestic infrastructure has been sufficiently hardened by IceFeather. Hesitant to hand over its golden goose, Moonshot AI offers API access instead of full weights. A lobotomized Kimi K4-Open gets released, while the real breakthroughs are monetized by selling tokens. Thanks to pressure from the Trump administration, OpenAI releases a cyber-lobotomized open model on par — which finally convinces the CAC that Moonshot should be allowed to let the world in.

The registry

Li Qiang, China’s Number Two, wasn’t kidding. What Beijing really cares about is “upholding the bottom line of AI security” — not the platitudes in Xi’s WAIC speech about “openness, win-win” and “mutual learning between civilizations” that the credulous foreigners ate up. Behind the scenes, Beijing has maintained deep insight into exactly how capable China’s top models are, and has been preparing for the moment it would have to step in.

The CAC publicly updates its AI model registry regulations. Whereas the rules once governed only models and algorithms able to “shape public opinion,” they now require companies to report any major update, self-assess for potentially catastrophic cyber risks, and give the government exclusive access, indefinitely. Remembering the fates of DiDi and Ant Group, every company knows better than to not comply.

First dibs

An internally assembled testing capacity — run through the MSS and the PLA Cyberspace Force — means that when Kimi K4 proves genuinely exceptional, Zhongnanhai is the first to know. The model is recognized as a potential “assassin’s mace” of the kind Xi directed the state to forge a decade ago,³ and there is no way the security state will allow it to be open-sourced. The Foreign Ministry protests that a global release would improve China’s standing; MIIT officials salivate at the domestic productivity boost. The security state is not having it. Open-sourcing would forfeit a cyber window of opportunity — and hand separatists and Falun Gong activists with some Azure credits the same hacking capabilities as the domestic security apparatus.

The summons

Two days after Kimi K4 is sent over to the state, Yang Zhilin and company are summoned to the CAC. Over long hours and litres of tea, a deal is struck. The “real” Kimi K4 goes to government entities — deployed on-premise into core governmental functions that benefit from such capabilities, including defense. Moonshot AI will release a lobotomized public version, tested by both company and government until its cyber capabilities are satisfactorily weakened. The lab gets a contract with lucrative terms; the state agrees to help “persuade” other Chinese model makers to sign away their compute, on the mainland and at Southeast Asian neoclouds alike. In a world where cyber capabilities threaten to become democratized, Beijing makes sure the tools in its own arsenal are of a different tier.

The leak

Parts of the meeting leak to Reuters — but no matter. At first, no one is surprised that the Chinese government gets better AI than the masses. Moonshot AI is now the proper national champion — and there is no actual proof it has made something transformative.

The bazooka points south

Then the attacks start. Beijing knows the Americans, armed with Mythos-level defenses for months, are watching for any sign of trouble — and it does not want to re-inflame the bilateral relationship when it cannot expect to extract meaningful concessions. Instead, it focuses its new cyber bazooka on the less prepared: wavering capitals in the Global South, less-developed neighbors, and eventually Europe. It is easy enough to hack into Paraguay’s citizen records; by early 2027, Asunción is frightened into dropping recognition of Taiwan. European technology companies in the semiconductor supply chain begin reporting cyberattacks at a much-increased scale. Brussels bureaucrats begin fretting about blackmail of uncertain origin.

The champion, insulated

Other Chinese labs, having signed over a good amount of their compute to Moonshot, progress slowly on their own capabilities — and are still barred by the CAC from releasing their weights. Collaboration between the government and Moonshot AI reaches new heights: building on Kimi K4, the lab now develops proprietary models tailored to government wants, and new-generation cyber weapons. On the outside, Moonshot AI still looks like a normal lab — consumer models, open-source research, AI products. The to-G side is heavily insulated: researchers are thoroughly vetted before being deployed to serve the government, and are required to surrender their passports. The CAC occasionally poaches star researchers from other labs under the guise of “national development consultation.”