The Largest Open Model Ever Just Went Free: Kimi K3's 2.8-Trillion-Parameter Weights Are Live

At 00:00 UTC on July 27, Moonshot AI did the thing it promised and plenty of people doubted: it put the full weights of Kimi K3 on Hugging Face, free to download. At 2.8 trillion parameters, it is the largest open-weight model ever released — and it lands three days after a coalition of American companies published a letter arguing that exactly this kind of release makes the world safer.

When we covered K3's launch on July 19, we flagged that the weights weren't actually out: the model was "open-weight in promise more than in practice," usable only through the API, with a download promised by July 27. The promise was kept, to the hour.

What actually shipped

Spec	Detail
Total parameters	2.8 trillion — the largest open-weight release to date
Architecture	Mixture-of-Experts: 896 experts, only 16 active per token
Precision	MXFP4 weights with MXFP8 activations, via quantization-aware training
Context	1M tokens
License	Apache 2.0 — commercial use permitted
Availability	Hugging Face, from 00:00 UTC July 27, 2026

The precision detail is the interesting engineering story. Moonshot didn't train at high precision and then squeeze the model down afterwards — it used quantization-aware training, so K3 was designed from the start to be served in MXFP4. That avoids the accuracy loss that usually comes with post-training quantization, and it is why a 2.8-trillion-parameter model is servable at all.

One thing to be careful about: reported download sizes vary widely, from roughly 594GB to about 1.4TB depending on whether the figure counts raw weights or adds runtime overhead and KV cache. Anyone planning to self-host should size against the larger number, not the smaller.

Eight days that framed the open-vs-closed fight

The timing here is remarkable, and it's worth laying out because each event changes how the next one reads.

Date	What happened
July 19	K3 announced: tops Arena.ai's Frontend Code leaderboard at 1,679 points, ahead of Claude Fable 5. Weights promised for July 27.
~July 22	White House OSTP director Michael Kratsios accuses Moonshot of an industrial-scale distillation operation against an Anthropic model to build K3.
July 24	Jensen Huang's first X post shares an open-weights letter — 25 signatories, defending open models and the legitimacy of distillation.
July 25–26	The letter roughly doubles to 50 signatories. Google signs. OpenAI quietly signs. Anthropic doesn't.
July 27	The weights drop. The largest open model in history is now a free download — from the lab at the centre of the distillation accusation.

Both sides of that argument just got evidence in the same week. If your position is that open weights spread capability, lower costs and reduce dependence on a handful of American vendors, K3 is the strongest example yet. If your position is that releasing frontier-class weights is irreversible and unenforceable, K3 is also the strongest example yet — because there is now no mechanism by which anyone, in Washington or Beijing, can un-publish it.

The self-hosting argument

There's a practical wrinkle that cuts against the usual framing. Using K3 through Moonshot's API means sending your prompts to a Chinese company's servers — a real problem for a lot of Western enterprises and a reason many wouldn't touch it. Downloading the weights removes that objection entirely: run it on your own hardware and no data leaves your infrastructure.

So the open release is simultaneously the thing US policymakers are most nervous about and the thing that makes a Chinese model safe for a Western company to actually use. That's the tension the open-weights letter was gesturing at with its sovereignty argument, arriving as a concrete case three days later.

The caveat is that "free" isn't cheap. A 2.8-trillion-parameter model — even sparse, even at MXFP4 — needs serious multi-GPU infrastructure. In practice this is a release for clouds, well-funded labs and large enterprises, not for a workstation under a desk. Most people will still reach it through somebody else's API.