Kimi K3 Will Be Dethroned

Kimiko 3 has been dethroned. How? By Quinn 3.8 max and all 2.4 trillion parameters of it. Command AI But this is balls to the wall parameters. Um It's What? two It's balls to the wall? You never heard of balls to the wall? [laughter] No. Everything thrown at it including the kitchen sink. 2.4 trillion parameters 95 act billion active parameters when it's being used cuz it's a mixture of expert model. And it's another one going toe-to-toe with I think this is a fable and they're going to be releasing the first time we will release the weights of a max class model. And they're going to be released next week. I think this was this week so 20 third. So yeah, still a few more days left. But yeah, 2.4 trillion parameters and they asked it to build a self-evolving harness. So basically they asked it to create all my CLI [music] and over 10 days long horizon autonomous coding run built this harness. And I'm guessing that this is a coding harness or or even command code similar thing to that. It built that. Um under a loop engineering setup. Kimiko 3 was like an amazing model, but it had a super like it was a lot of parameters. I actually think it's bigger than this one. Kimiko is 2.8 billion. Whereas this is 2. 4 billion. It really does come down to the more parameters you throw at something, the better the model is going to be. So and Quinn have a series of open source models, right? They do have like a 12 billion. In fact, here. Basically, this was a reproduced a research paper and improved it and it wrote a bunch of code to be able to do that. And what it did, it fine-tuned that 8 billion parameter model to be able to improve the paper, basically. So, it tuned its own model. Well, it tuned another model to be able to to do it. So, they have all of these smaller models that you can do more fine-tuning with and this and that. But again, this is the balls to the wall model where they've just thrown everything at it. We're still waiting on the weights. They've done the same thing as Kimmy where they're like, um, you know, the we'll release you the the sort of closed source version first and then we'll release the the open weights. So, for now, you can run it in Quen Cloud, which uh, I don't I don't know if you're interested in this sort of stuff, but Alibaba had a token plan, which was $10 for Quen, um, MiniMax, GLM, DeepSeek. They had a really, really good deal. They've recently increased the price and actually stopped signing up to it and then they dropped this Quen Cloud plan, which check this. It they've kind of replaced it. Multiple models, one plan, where you get Quen, I don't know if you can see that, Quen, GLM, DeepSeek, and Wan, which is an image generation model, for $6 a month. So, I'll put a link down below, which I think gives you a discount or something. It will be a referral link. I'll I'll drop that somewhere. Um, for $6 a month for all of that. I doubt it would last you very long and it looks like they're already trying to they're already shutting down some of the Oh, it's the Maybe no, maybe I think that's as good. I think this is a good thing. It's the it's the it's the 5-hourly limit. They're just giving you a week to So, that's really good because sometimes you're on really heavy coding days, sometimes you're not. And you know, to be restricted by those 5-hour limits can be really frustrating. So, uh, this gives you more autonomy in in balancing how you like to work by only limiting you every 7 days. There there is a downside, though. If you prompt something wrong or maybe something goes wrong and the model just gets crazy about trying to get what you want to do, but it can't do it and it just gets in a self loop trying to do it, you can potentially run through your whole week uh week's worth of uh credits in one prompt. Like, this could happen. Yeah, it's difficult, man, because it gives you autonomy. Maybe there should be an option to like turn on the 5-hour limit or not. It It should be like there should be a limit where you can set like if it hits like 20% stop the agent or like pause it. I don't know if it can how that technically would work because your with your chat and with the history and everything, I don't know how that would work, like hitting a pause. Mhm. Don't know, but Anyway, but it it's cool that these Chinese models are getting this capable and as you were like showing me these, I was thinking, like, do we need something significantly more powerful than Fable 5? Like, Fable is awesome, right? But, do we need anything way more powerful than Fable? I mean, we'll take it, obviously, but like realistically What can it do? Mhm. I mean, it's just better like Fable will still make mistakes. Fable will still hallucinate. Fable will still um you know, yeah, hallucinate. I think it's the worst one there, really. So, [snorts] I think it's just better is not necessarily more intelligent, but better could be maybe it's faster. Maybe you talk about it being faster. Maybe you talk about it being less hallucinogenic or secret fantasy as you to it. isn't it like the the model being faster based on like the servers and the the hardware that it Yes and no. access to? Yes and no. Like there's only so much physics that you can throw at a model to make it faster and I think Yeah, I think maybe that maybe someone will come up with something. Maybe it's the architecture itself like we talk about um you know, obviously mixture of experts was something that wasn't available in first instance, but we found out that you can have these massive massive models, but you can defer the task to um uh Google released Gemma released two models that this is their open source models called E2B and E4B and they did a little video about what the E means. And it's an it's a a method of embedding. It's called E is E is for embedding, right? And it's just so that they can have the same token exist on many layers. They embed the same token on many layers. And again, this is all way beyond our pay grade, but like the idea is that people are coming up with interesting ways to handle and deal with AI, right? So yeah. Speed can absolutely be achieved through that might be the same intelligence, the base intelligence, but the method in which it retrieves that intelligence could be improved on. It's 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. Command AI