On 27 July 2026 at midnight UTC, Moonshot AI released the full weights of Kimi K3, its 2.8-trillion-parameter model. The promise was kept on the announced date, which was not guaranteed. The file weighs around 1.4 terabytes, making it the largest open model ever released, and puts self-hosting out of reach for the vast majority of users.
We have covered Kimi K3 twice: first at its announcement, noting that the weights were not yet available, then about its appetite for reasoning tokens. Here is the epilogue to the first story, and it is instructive.
The promise is kept
At the model's announcement in mid-July, we flagged an important nuance: Moonshot presented Kimi K3 as open, but the weights were not downloadable, with a release promised for 27 July. We wrote that a promise on a calendar is not a file on your hard drive.
The file arrived, on time. The full weights were made freely available at midnight UTC on 27 July. That deserves noting, because the industry has accustomed us to openness announcements being postponed indefinitely. Kimi K3 officially becomes the largest open-weights model ever published, and it arrives amid increasingly pronounced Sino-American competition over openness.
The figure that puts things in perspective
Now the material reality. The weights take up around 1.4 terabytes, using a compression format called MXFP4.
To grasp what that means, a useful reminder. Quantization, or compression, involves storing each parameter with less precision, a bit like going from a lossless audio file to an MP3. The model then takes up far less space, with a quality loss that is often moderate. And despite this aggressive compression, we are still talking about 1.4 terabytes.
In other words: the file alone would not fit on most laptop drives, and running it requires far more than storing it, since as we explained in our article on Mixture of Experts, all parameters must remain available even if only 1.8% are active at any given moment. Moonshot recommends serving the model on clusters of dozens of accelerators linked at very high bandwidth. In practice, most teams will access it via inference providers rather than hosting it themselves.
This is exactly the distinction we detailed in our guide to local AI. Open means that anyone has the right to download, modify and deploy the model: that is total legal freedom, and it has real value. It does not mean that anyone has the material means to do so. For an individual, genuine democratisation remains on the side of open models with 3 to 30 billion parameters, not 2.8-trillion monsters.
So, what is this release for?
The question is legitimate if almost no one can run the model at home. The answer comes down to three points.
Scale-level sovereignty. A state, a large corporation, a university with a computing centre can now host a top-tier model without depending on any foreign API, and without anyone being able to take it away from them. This is precisely the issue we described in our article on open weights, and it concerns institutions, not individuals.
Research. Available weights can be inspected, analysed, modified, studied. That is irreplaceable for understanding how these systems work, and it benefits the entire scientific community, including those working on AI safety.
Competitive pressure. A freely available model that rivals the best closed models puts permanent pressure on prices. This dynamic partly explains why Google is cutting its rates and why Anthropic is positioning Opus 5 on value for money. You benefit from it even if you never touch Kimi K3.
What to take away
This release is a real milestone, and not just a symbolic one. It confirms that the Chinese openness strategy is not a communications posture: when a date is announced, it is kept, and the file is genuinely there.
But it also reminds us that the word open deserves to be read with precision. A terabyte and a half of freely downloadable weights is immense freedom for institutions and an abstraction for most individuals. Both observations are true at the same time, and it is this nuance that is most often missing from discussions about open source in AI.