Why does everyone from Wall Street to the White House care so much about Kimi K3?
Earlier this month, when Moonshot, a Chinese AI developer, released its Kimi K3 model , it quickly became a lead story in technology and geopolitics. On Monday, as promised, Moonshot released the model weights for Kimi K3 and working versions of it became available on model hubs hours later.
I’m writing this post to offer a sketch of what all the fuss is about and to follow up on an opinion piece I had in the Globe and Mail last week about open models and the future of AI.
My editorial, titled ‘Canada’s new data centres could be obsolete before they open ,’ was about how smaller open-weight models are improving much faster than expected, and how, soon, more of us will be running AI on our own devices, relying less on AI in the cloud and thus less on large data centres.
The Kimi K3 release is part of this story but points to what’s happening at the higher end of AI development.
My brief snapshot is this:
The most advanced language models in the world are the frontier models coming out of labs at OpenAI and Anthropic, currently GPT-5.6 and Opus 5. They’re closed (companies don’t make them free to download) and accessible only in the cloud, where use is metered by the token. Companies that make heavy use of them are spending hundreds of thousands, even millions, of dollars a month in token fees, a price they’re willing to pay to have access to the very best models.
Alongside the development of frontier models from a cluster of companies out of Silicon Valley are various smaller start-ups and developers around the world, most notably in China, making open-weight models. They post these models on model hubs, free to download and use, either on your own system if they’re relatively small, or by paying a cloud service, like the ones Amazon or Microsoft own, to host and serve them for you if they’re large.
Using an open-weight model on your own computer or system, offline, gives you access to completely private AI and unmetered use of it. But it will only be as good as the size of the model you’re able to run (more on that below).
Downloading a larger open-weight model, like Kimi K3, and, if need be, paying a cloud provider to host it gives you more privacy (though not complete) and the chance to tailor the model to your use cases, through post-training on your own data or by connecting it to a special database (for law, business, etc.).
The big deal being made about K3 is that, on some benchmarks , it now ranks as the most capable language model in the world, exceeding even the best from OpenAI and Anthropic. On others , it’s a close second or third. And as you may recall, the US government forced Anthropic to pull Fable, one of its most capable models, for about two and a half weeks in June, fearing it could be used to find and exploit software vulnerabilities.
So K3 is comparably powerful, but free to anyone who wants it, and without any guardrails.
How did Moonshot make K3? How do the other companies producing open-weight models make them? And why?
How to make an open-weight model
Developers make these models through a long-standing process in machine learning called distillation. One site defines this as a technique where
a smaller model (known as the student) learns to mimic the output of a larger, more complex model (the teacher). The goal is for the student model to achieve a level of performance close to the teacher, but with significantly fewer parameters, making it more efficient for deployment.
It’s controversial because Moonshot and other companies use normal commercial accounts to access frontier models for this training, in violation of those providers’ terms of service. OpenAI and Anthropic make their models available to customers to use but not to clone. And that is what distillation involves, though it doesn’t quite result in a perfect clone. More like an adequate substitute for many purposes.
As the Verge points out , with few exceptions, open-weight models are not open source. The weights are the parameters acquired in training; they do not reveal things like the training data, model architecture, or configuration methods. Some have speculated that Chinese models could be “poisoned,” with their weights containing a backdoor opened when “an unknown trigger appears,” making them unlikely candidates for use in defence or critical infrastructure outside China.
What open-weight models mean to the future of AI
A few broad takeaways follow from Kimi K3 and the larger trend it signifies.
The time lag between the best frontier models being released and the best open-weight models catching up to them in capability is closing from a year to six months, to now four months. This is likely to continue, turning frontier AI into a mere commodity.
Although the largest frontier models are soon followed by open-weight models with similar capability, both are still quite large. Not something you can run on cheap hardware.
But this too is rapidly changing. A mid-sized firm or institution can now run one of the biggest open-weight models, like K3 (at 2.8 trillion parameters), in a quantized or light-sized version on a cluster of consumer machines costing in the low tens of thousands, at a fraction of a datacentre’s energy draw.
Apple is at the forefront of making this possible. Its shift to computers that combine memory and processing power on a single chip allows people to use its devices — or stacks of them linked together — to run sizable models, once again at a relatively low energy draw.
To lend a sense, as one commentator wrote earlier this month:
Apple shipped a Mac Studio you could configure with 512GB of it, for about $9,500, drawing a couple hundred watts at the wall. …DeepSeek’s 671-billion-parameter V3 runs at 20+ tokens per second on a single 512GB Mac Studio. Kimi K2 Thinking — a trillion-parameter model — runs at ~15 tokens per second on two of them. GLM-5.2, one of this summer’s frontier-class releases, fits on one box at 4-bit. A machine that plugs into a normal wall outlet, makes less noise than your refrigerator, and runs a model that would have been the best AI on Earth eighteen months ago — with the door closed and the internet cable unplugged.
He also added that “When Kimi K3’s weights land on July 27 at a projected ~650GB–1TB quantized, no single consumer box on the planet will hold it. A few tethered Mac Studios will.” (This has now been confirmed .)
Seeing where things are headed, Nvidia is following Apple’s lead and has begun selling a small computer with an integrated memory chip, its own version of the Mac Studio.
The political economy of open-weight models
In the wake of what Moonshot did with K3 this month, commentators like Ben Thompson and Dean Ball foresee frontier model makers and the US government raising greater impediments to distillation by Chinese firms. Not enough to end it, but enough to make it more costly and to slow it down.
There are differing opinions as to why American companies like Meta or Google, or why China itself, would support the creation and distribution of open-weight models. One view is that it resists market capture by one or two AI providers at the frontier, or by American companies alone, allowing a larger ecosystem to emerge around model development.
Last week, a number of large tech companies, including Nvidia, OpenAI, and Microsoft (but notably, not Anthropic) signed an open letter in support of open-weight models. The letter touts three main benefits of these models. They give users greater control and access to AI, since they can be modified and run on your own equipment. They avoid lock-in to one provider. And they’re more efficient, in economic and environmental terms, “match[ing] the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems.”
The letter also concedes a key danger of open-weight models: they’re beyond the control of the original developer and can be modified and used for cyberattacks or other nefarious purposes. But the companies here believe the very same models can make it easier for defenders to acquire “comparable capabilities so they can detect, simulate, and respond to emerging threats.”
Anthropic responded on Monday in a post noting that it isn’t opposed to open-weight models, but thinks the US should not sell its most powerful chips to China, should “crack down on industrial-scale distillation operations,” and should impose mandatory safety testing on the more capable models, open or closed.
But Anthropic may be on shaky ground calling for a crackdown on distillation, since it isn’t clear that model creators have intellectual property over the process of distillation. Ben Thompson suggests the US should pass a law banning frontier model makers like OpenAI and Anthropic from prohibiting distillation, so that American and other Western companies don’t have to rely primarily on Chinese companies to obtain open-weight models.
Regardless of how these models are made, it seems likely that more of our AI use will migrate to open models running on smaller equipment or with less reliance on the cloud, potentially making it cheaper, more private, and less harmful to the environment. ■
{To receive new posts by email, follow on Substack .}