The model ships with open weights and the most permissive licence anyone hands out. You cannot run it at home, but any provider can stand it up and sell it to you.
- DeepSeek-V4-Pro-0813 lands on 13 August with 1650 billion parameters.
- The licence is MIT, which in practice puts no limits on use.
- The weights are in FP8, which halves the space they take.
The number I am watching is not a trillion and a half. It is three letters: MIT.
That is the licence that says do whatever you like, just leave our name somewhere. No user threshold, no ban on commercial use, no clause about competitors.
Who can actually use it
Not you, if you were thinking of standing it up on your own machine. One thousand six hundred and fifty billion parameters want a hall, even at eight bits per number.
The licence, though, is written for the middlemen. Any compute provider can stand it up and sell it by the hour, asking nobody and sharing no revenue. From there the model reaches you at a price somebody else worked out.
And that is exactly the move. Hand out weights with a licence that gets in nobody's way, and a dozen providers will have it up within a week, so your model is suddenly everywhere. The closed competitor has to build the same thing alone.
The eight bits
FP8 means every number takes half the space. The files are smaller, the memory stretches further, the speed goes up. You pay in a little accuracy, which at this size is barely felt.
So they are shipping it in the form that is cheapest to stand up. That is not an accidental decision; it is part of the same calculation.
For you today: look at which provider it comes through and at what price. With a licence like this there is usually more than one option, and the gap between them is not small.