Open Source Will Change the Game (Kimi K2)

Open source AI is going to make all these AI companies obsolete. Kimmy K2. So this is open weight model. Yeah. An open source model with uh a trillion total parameters. We're at the moment we're talking about billions like 500 billion like whatever. This is a trillion parameters and achieve state-of-the-art performance uh frontier knowledge in math coding among other non-thinking models. Now this is quite important because it's non-thinking. There are two models. It's Kimmy K2 base and Kimmy K2 instruct. Obviously, this is the foundation model in which this instruct has been been trained on. And it is performing quite well compared to many of the other ones. I mean, we're just behind the industry leader, which is uh Claude here. Claude being 72 whereas Kim K2 is 65 on the Sweetbench. Multilingual 51 versus 47. Live Codebench V6 is actually beating Anthropic here. I mean, yeah, it's incredible. And this is from this is a Chinese model, so it's uh something to take note of. And it's a very cheap model. So, China are really pushing the pushing the envelope when it comes to creating really really good. Can you download this model and run it locally? In theory, you can. How many gigabytes is that? and how many because one trillion is like crazy. That's pretty big. I think this is just going to shake up the industry again. I used it for a bit and it was really slow and then I watched some other person's video on it and they were getting 200 tokens a second and I think Opus was something like 40 or 60 tokens a second and I was like well why you know I the thing is about Deep Seek. Deep Seek came out and whilst it wasn't necessarily the best, they've got a lot better now by the way. Uh whilst it wasn't necessarily the best, it just it was just a point in time where it was like China just came out with this super cheap, very capable model against all of these closed source options. It wasn't necessarily a good model. That's not what the point is. I think this is the same again. This is the same a similar kind of shift in the the industry once again just to have another model that's non-thinking doing really really and scoring really really well. This question existed kind of like from the start of this AI hype that uh if open-source AI is going to make all these AI companies obsolete. uh and this question comes up every time when China especially releases an open-source model that is you know on par with these like expensive models that have hundreds of billions of dollars in investment and there is an open- source one that you can run locally by the way I searched it's about just over one terabytes that means you need uh a computer with about one terabytes of VRAM I it's the graphic card RAM that you need which is like insane with your typical PC uh setups you will not get that you get like 20 gigabytes of VR RAM on a high-end computer a terabyte we are talking so you need quite a few uh Apple Mac uh pro thingies to get to that level so that is insane that is an insane size and it's crazy that Now we can have these models as an open-source model that we could theoretically run on our own hardware or maybe a server that you rent that you know can do that and then the question becomes you know what are these AI companies uh doing what is the race this race that is going on for because if open source is catching up constantly how can these models win how can chedd win if you could just use this model for free or for the fraction of the cost. What justifies Open AI's what trillions of you know planned um um you know uh investment. It's insane. I I I don't know. This is the comparison here. So we've got the open- source models here and then you got the propriety models here. So 53 on live codebench which beats out everyone. The winner here is that uh Opus 4 is is beating them on multiple uh multiple multiple. Um you can kind of see that, you know, they they they have they are doing very well. They're doing very well against the closed course or source or proprietary software here. But you're right, they are catching up. There's no Grock 4 there. I don't, you know, unless you go through all of these and and look at the numbers that Grock 4 reported, there's no Grock 4 against some of these. So, it'll be interesting to see how it compares to Grock 4. This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter.