Moonshot gives away the strongest open AI model in the world.

Since 27 July the strongest freely available AI model in the world has been sitting there for download.

Moonshot AIChinese AI company. It develops the Kimi models and releases them for download, rather than only selling access to them. published the full weightsThe numbers a model learned during training. Whoever holds them can run the model themselves, without asking the maker. of Kimi K3 on Hugging FaceA platform where developers upload and download AI models. For open models it is the usual address. View the model page. 2.8 trillion parametersThe individual dials inside the model. More parameters mean more compute, not necessarily better answers., 104 billion of them active per tokenThe unit a model breaks text into. A token is roughly half a word to a whole word.. A context windowHow much text a model holds in view at once, question and answer together. Once it is full, the beginning drops out. of just over a million tokens.

On the Artificial Analysis Intelligence IndexAn independent comparison test that folds several exams into one score. It measures only what it tests, and is read as a ranking anyway. See the index Kimi K3 reaches 57 points. That is the best score an open model has ever achieved. The next open model sits at 51.

After roughly a day it had 99,200 downloads.

The interesting number is a different one. The weights come in 96 files totalling around 1.56 terabytes.

1.56 terabytes fit on no work machine

That is more than the hard drive of most business computers holds.

It is also not a file you download and start. It has to sit in memory while computing, and in specialised hardware at that.

Anyone wanting to run this model needs a data centre or rents one. Which puts them back where they started: With a provider selling them compute.

Anyone can download it. Very few can run it.

The objection specialists will raise

The 1.56 terabytes are the total size. During computation only a fraction is active at once, 104 billion of the 2.8 trillion parameters per token.

Anyone who knows the field will object that my argument about running it is too crude. On top of that come methods that reduce memory needs further.

The objection is right and it still does not change the result. Even a fraction of 2.8 trillion parameters is beyond what a business without its own server hardware can handle.

I expect this to shift over the coming years. In July 2026 it is what it is.

The licence is not an MIT licence

An open licence sounds like take it and do what you want.

Kimi K3 comes under a licence of its own. Use, modification, distribution and commercial sale are permitted.

But: Anyone offering a model as a service and turning over more than twenty million dollars in twelve months needs a separate agreement. Products with more than a hundred million monthly users have to credit Kimi K3 visibly.

For a normal business neither threshold matters. The thresholds still show something: This is not a gift, it is a business model with an allowance.

Open weights mainly benefit those who sell compute

The word open suggests you are making yourself independent of a provider. With a model this size that is not what happens.

On the same day Kimi K3 went free, Nvidia, Microsoft, Meta and PalantirUS company for data analysis software, mainly for government and defence. It does not build large language models of its own. signed an industry letter backing open model weights. Google and OpenAI followed over the weekend. The providers of closed models are the ones backing open ones.

I do not believe anyone changed their convictions here. The larger the freely available models get, the more expensive running them becomes. And the more surely everyone ends up back with a provider.

That does not make the release a bad thing. It is good for research, for verifiability and for keeping prices under pressure. It is simply not the declaration of independence it is being celebrated as.