Hook
More breakout videos from this creator.
Everyone thinks you can't run Kimi K3 on a laptop. The 2.78 billion parameter model with almost an entire terabyte of model weights, but they're wrong! And I'm gonna tell you exactly how to run this on your laptop. Hit that save. Let's get in. Weight aware streaming tensor engine. That's right, it's called waste. And you can see here from the GitHub stars that almost nobody knows about this, so we are going to get it. This is a clean no dependency C program made by Marco Bambini. Thank you, Marco Bambini. It runs the full weight Kimi K3 on a high-end consumer laptop. It does need 64GB of RAM, but it works! A good way to think about the way that it actually works is pretend you have like 80,000 experts that are all perfect and really, really good at knowing everything in their field. Okay, and then when you ask a question, this pulls 4% of them, wakes them up, asks them the question, puts them back to sleep. It actually works very similar to what we were talking about before with the ESP32 where everything stays on the SSD and only the necessary things move over to the fast VRAM. This is a major upgrade to architecture. And as these architecture upgrades continue, we are going to be able to get better and better models on simpler and simpler PCs. Yes, there will be bottlenecks, yes, of course there will be restrictions, but the future is now. The time to take advantage of this is right now.