Moonshot put a 2.8 trillion parameter model on Hugging Face, and the download alone is 1.56 TB

Good day,

Yesterday I wrote about a model called Small, 594 billion parameters, and you still cannot download it. Today I have the opposite story.

Moonshot put Kimi K3.1 Pro on Hugging Face on November 4, weights and all. The model is huge, 2.8 trillion parameters in total, with 104 billion active for each token. And this time the download link is real.

Real does not mean light. The weights come in 96 files and take about 1.56 TB. That is not a laptop download. Moonshot points people to vLLM, SGLang and TokenSpeed, and you need a rack of fast GPUs just to hold the thing. A model you can download is still a model most of us will only meet through the API, at 2 dollars in and 10 dollars out per million tokens.

Then there is the license, and this is the part I liked least. It is Moonshot's own Kimi K3.1 Pro License, not a standard open license. If you sell the model as a service and make more than 20 million dollars in a year, you need a separate deal with Moonshot first. Very large products also have to show the model name. So the weights are open, but the rules around them are not small print you skip.

What about the scores? Moonshot says K3.1 Pro gets 65.7 percent on Terminal-Bench 4.0, next to 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. Those are Moonshot's own numbers, and nobody has checked the downloadable weights on their own yet. I am keeping the praise back until someone outside Moonshot runs the same tests.

Still, this week gave us both kinds of giant. One you cannot download, and one you can, if you have a server room and a license check first. Open is a real step. It just is not the same as easy.

Cover image: AI-generated