The Most Powerful Model You Can Download Is Chinese and Free

You can download a frontier model right now for nothing. It weighs 1.56 terabytes, it arrives in 96 shards, and it was built in Beijing.

Moonshot AI published the full weights of Kimi K3 on Hugging Face on July 27. Two point eight trillion parameters in a mixture-of-experts design, a one-million-token context window, native vision. Quartz confirmed it as the largest open-weight model ever released to the public.

Tom’s Hardware reported that K3 beat Claude Fable 5 in the Frontend Code Arena benchmark. Not matched. Beat.

Three days earlier, DeepSeek shipped V4 Pro as a stable release under an MIT license, weights available on day one.

So inside four days, two Chinese labs put roughly four and a half trillion parameters of frontier capability on the open internet and asked for no money.

Semiconductors sold off on both sides of the Pacific the same week. 24/7 Wall St. tied the Nasdaq’s chip slide to Chinese competitive pressure on July 27, and Japan’s Nikkei dropped more than 3 percent on a chip selloff the following session. Everyone remembered January 2025 out loud.

The architecture is what makes the size less absurd than it sounds. A mixture-of-experts model routes each request to a small subset of its parameters, so the thing is enormous on disk and much cheaper to run than the headline number implies. Tom’s Hardware framed the release as China engineering around US compute export limits, which is the correct frame. Restricting chips did not stop the model. It shaped the design.

The comparison to the first DeepSeek shock is doing a lot of work in the coverage, and it undersells what changed. In 2025 the story was price. A Chinese lab had matched a US model for a fraction of the training cost, and the market repriced the assumption that frontier capability required infinite capital.

This is a different claim. Moonshot is not undercutting anyone. It set the price at zero and handed over the artifact.

You cannot compete with free by being cheaper. You compete with free by being necessary, which means the entire US model business now rests on service, reliability, and distribution rather than on the weights themselves.

That is a survivable position. It is also a completely different business than the one investors were told they were funding, and the valuations were written when the weights were the asset.

Nathan Lambert called this an open-weights escalation rather than a release, and escalation is the right word. Moonshot did not have to ship 2.8 trillion parameters to be competitive. It shipped them to establish a ceiling that US labs now have to answer above, for free, in public.

The response from San Francisco so far has been a set of blog posts about responsible deployment.

The obvious objection is that free is not free. A 1.56-terabyte model does not run on your laptop or your startup’s credit card, and serving it at any real volume means renting a cluster that costs more per month than most seed rounds. Together AI and Modal shipped day-zero hosting, which tells you exactly how many people intend to self-host: not many.

The license is the sharper catch. K3 ships under Moonshot’s own terms rather than MIT, and the version reported by explainx gates commercial inference above twenty million dollars in annual revenue. Open weights and open source are not the same document, and the difference lives in a clause you accept by downloading.

TechTimes flagged an undisclosed hallucination-risk profile ahead of the release, and the alignment work behind a Chinese frontier model is not something anyone outside Moonshot can audit. Downloading the weights does not mean understanding them.

None of that changes the strategic picture, because the strategic picture was never about whether you personally can run it.

Every developer in Jakarta, Lagos, São Paulo, and Warsaw who builds on K3 is building on a Chinese architecture, with Chinese tooling conventions, against Chinese benchmarks. That is the whole play. Moonshot is not leaving licensing revenue on the table by accident. It is buying the default, the same way Android bought the phone market by costing manufacturers nothing.

US labs have spent two years explaining that their moat is the model. DeepSeek made the price argument in April and the industry called it a cost optimization. Moonshot has made the argument again with the actual weights, and the honest version of the response is that a moat you can download in 96 shards was a distribution advantage wearing a lab coat.

The thing to watch is not the benchmark table, which will be stale by September. It is whether a serious US lab answers this with open weights of its own, because right now the only companies giving away the frontier are the ones that were supposed to be behind, and the ones charging twenty-five dollars per million tokens are the ones explaining why that is still the responsible choice.

Sources: Quartz · Tom’s Hardware · Interconnects · MarkTechPost · TechTimes