This Is the Next Deepseek Moment (Ox Alpha)

A mysterious free reasoning model called Ox Alpha just appeared in Open Routter and it was scoring really high. For comparison, GBT Soul was scoring a 52. Fable was a 65. Ox Alpha 80%. Oh, wait. Just in we we know who it is. Command AI. So GLM 5.3 Flash is the is what we've previously referred to as OX Alpha and it is it's a doozy. The word flash should constantly be ringing in your head. I it it means like the smaller ones. The smaller ones. Exactly. So, highly competitive price, million token context window, which you know is a um is we've what we've come to as standard a 320 billion parameter uh mixture of expert model. This means there's 18 billion parameters MIT so open source open weights and previously ox alpha running entirely on Chinese AI chips. Oh wow. now. So 320 by comparison I think it was Kimmy K3 actually I think that was 2.8 trillion parameter model. This is a small model that's again scoring 80 on those uh those benchmarks which is just absolutely incredible. It was free by the way when it was Ox Alpha it was a free model. I don't know whether I made that super clear right but these are really cheap prices as a flash model. This is this is crazy. So, if we look at this here, it's a bit of a mess, but there's the GLM flash. The co this is the cost per task. So, it's all the way down here. This is where you want to be. Cheap, cheap, cheap, cheap. And the intelligence is all the way up here. There's soul. There's soul there. It's just nipping at the heels of soul there. Where's fable? I mean, fable's all the way over here. But if you look from a um intelligent perspective, what's GPT flash is around there. Can you see where my mouse is? Yeah. So, not far off of fable performance, but all the way down it for cost. And I think that's kind of all you need to know that this is I would describe this as a a deepseek moment. another deepseek moment where they're getting crazy performance out of you know very the good deep seek uh or like the deepseek moment that we are talking about was deepseek three two two uh maybe three actually I've forgotten it's yeah whatever the point is there on the chart we had deepse four that's why I'm saying yeah yeah yeah yeah deepse four is the latest one so deep sweep deepseek um this moment was they managed to train a very capable model off with like $2 million, which training a model on $2 million is is is unheard of at the time. It wasn't an amazing model, don't get me wrong, but it was it was it was unprecedented at the time. And uh to do it on Chinese hardware, to do it for that sort of price, it it's it's pretty pretty impressive. Now, Unsllo have already got it as GGUF, which isn't optimized for Mac, but it can run on Mac. And this is what we're looking at here from from a size perspective. All the the the full fat version is looking at about 200 gig. Um, but I can run, you see these ticks cuz I've put in my M5 Max. I could run the one bit version on my M5 Max. Uh, but it's still taking up 93 gig. So, it's still quite big. But it can run it can run on it. Um I look to I I'll wait and see if there are some LMLX versions. There are some MLX versions coming out which I'm tempted by. But you know on swap of I think day zero they had this ready day zero to be able to run it cuz the if we go here again I'm teaching you to suck eggs when it comes to local models here. But these are these are the weights. You can't do anything with these. you have to you have to convert it or quantize it or do whatever to actually get it working on sort of consumer hardware. So these are the weights that they release. There's not a lot you can do. And if we look at the files here, these are 328 gig the total size of all the all the these these are the weights. I've got this coding plan here. They've actually deprecated this, you know, these prices here. By the way, this is monthly. This is quarterly. This is yearly, so obviously save the more you buy. But they've updated the plans here to be credit based. Now, honestly, I don't know how they did this previously. Surely they would, you know, surely they're all credits based. Like I had this one is five times the usage. Like did they just abstract the number of tokens you used and it was based on I'm not too sure. But I think this new plan gives you a clearer picture on how many tokens you're actually using because your credits, you get 10,000 credits a week on the light plan as an example. Six times that you get 60,000 credits. 14 times that, you get what's that? 140,000 credits. Um, so they've updated the plans. I think you can I'll put my link down below, by the way, if you want to. In fact, let me give you the link because I give you a discount basically if you were using it. I put it in the chat. Oh, sorry. Last thing to say, it's multimodal. The previous version wasn't multimodal. So, this can look at images and actually review the work that it's done. So, really really impressive model from ZAI there. Um, the current the new this week's new best model from ZI. This week's Yep. 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. Command AI.