AINEWS
Moonshot AI's Kimi K3 publishes its model for free, but loading it requires 16 high end Nvidia chips, and running it as designed costs roughly $2.4 million in hardware, plus factory grade power and cooling.
Imagine a three-star restaurant publishing its signature recipe for free, then telling you the kitchen costs about 20 million yuan. That is roughly the situation behind the latest wave of "open source" AI models, and Moonshot AI's new Kimi K3 is the clearest case yet.
K3 is genuinely free to download. Moonshot, one of China's better-funded AI labs, has released the model's trained parameters, the numerical settings that determine how it responds, so anyone can fetch them and run the model themselves. The catch is what "running" means. According to a Chinese explainer translated in ChinAI #369, merely loading K3 into memory requires at least 16 of Nvidia's H200 accelerators, the current top-of-the-line chips for AI training and inference. Running the model the way Moonshot recommends takes 64 cards wired together as a "super-node," a configuration that costs roughly 17 million yuan, about $2.4 million at current exchange rates, before you pay for the building to put it in.
The Chinese explainer frames the gap with the kitchen analogy. A three-star Michelin recipe, given away, still requires 64 professional-grade stoves, factory-grade power, and the right kind of ventilation. So does K3.
A fully loaded K3 setup draws 45 kilowatts. A typical US household is wired for about 20 to 25 kilowatts of continuous draw before the main breaker starts complaining. Chinese apartments are not built for that kind of load either. K3 is not a model you run on a desk. It is a model you run in a room that already has three-phase power and an industrial cooling system. The Chinese explainer puts the hardware floor for K3 well above what an individual or small office can realistically supply.
This is a sharp break from K3's predecessor. K2, an earlier Moonshot release, was a compressed build that enthusiasts actually got working on a Mac Studio, a desktop computer that costs about as much as a used car. K3 is positioned by Moonshot as a different species, a data-center or enterprise-only model that happens to ship with public weights. The "open" label still attaches to the file you download. The "free" label still attaches to the price. Neither label describes the cost of doing anything with the file once you have it.
Most "open" AI releases, K3 included, do not include the source code for the training pipeline, the data filtering, the evaluation harness, or the tooling that turned raw compute into a working model. What they publish are the weights, the final state of a model after training, frozen at one point in time. In ordinary software, open source means the source code is published: anyone can read it, modify it, recompile it, and ship a new version. With a weights-only release, you can run the model. You can fine-tune it on new data. You cannot, in any practical sense, rebuild it from scratch or inspect why it behaves the way it does.
Treating a weights-only release as the equivalent of open source lets the marketing language outrun the operational reality. The labeling carries specific expectations in software procurement, regulation, and security review. A model that publishes its weights is closer to a compiled binary than to source code. The Chinese article's point, gently, is that a small developer whose boss says "just run K3" is being asked to spend a year's salary on hardware to load a file that arrived in an email.
K3 is not unique in this. Meta's Llama family, Mistral's larger releases, and DeepSeek's recent flagship models all ship weights without the full training stack. The "open source" framing has become a standard industry move, both in the United States and in China, because it is technically defensible and commercially flattering. What is new with K3 is how explicit the hardware ask has become. Sixteen H200s to load. Sixty-four to run well. Forty-five kilowatts. Seventeen million yuan. Those numbers are not hidden in a footnote; they are the headline of the deployment story.
For anyone evaluating K3, or any of its open-weights peers, the practical question is not whether the model can be downloaded. The question is whether the rest of the stack is in place: the accelerators, the power, the cooling, the engineering time to wire it all up, and a workload that justifies the spend. The model is free in the way a recipe is free. The kitchen, in this case, is the data center, and the data center is the part that costs.