On 27 July Moonshot AI uploaded the full weights for Kimi K3 - 2.8 trillion parameters, the largest openly available model to date. We wrote about its debut ten days ago. Now anyone can download it and run it on their own machine. In theory. In practice the weights weigh in at around 1.5 terabytes and need hardware that few people have.
- The full weights for Kimi K3 have been published on Hugging Face - 96 files, around 1.5 terabytes, under its own 'Kimi K3 License.'
- According to Moonshot: 2.8 trillion parameters, of which 104 billion are active at each step; context up to 1 million tokens; multimodal.
- The model debuted on 16 July; the open download went live on the 27th.
You download 96 files, install them on your own machine - and nobody can stop or switch it off anymore. That distinction is exactly what gets lost in the news. 'Open-weight model' doesn't mean free access through someone else's website. The weights are the actual numbers that training arrives at - the billions of settings that make the model what it is. It sits with you, on your own server, with no outside middleman in between.
What's new, given that we already wrote about K3? The access, and the bar. Closed labs keep their strongest work behind an API and a subscription. Moonshot puts a model of the same class on the table - here it is, download it, it's yours. But let's not kid ourselves: open and accessible aren't the same thing. Weights that weigh a terabyte and a half need machines that an ordinary person, and most companies, simply don't have. The real audience is the clouds and providers who'll run it and sell cheap access to it.
The real effect is the pressure. Every time an open model gets close to the top, the price of the closed ones drops - we've already seen it happen a few times this year. Whoever builds on AI today will pay less tomorrow for the same intelligence - precisely because of releases like this one. That's why we watch them closely, even without a terabyte and a half of free space. And without a server for them.