Your PC can now run a 2.78T AI model. And it only needs around 8GB of RAM.
That sounds impossible.
But an open-source project called kimi-k3-in- c is making it work on a regular PC. No GPU required.
The engine streams model experts from disk and only activates 16 out of 896 experts for each token.
It still needs a fast, large NVMe drive and isn't exactly speedy.
But the bigger idea is fascinating:
You don't always need the whole model in memory to use it.