Cancel your subscriptions, Ox-Alpha is here! (GLM 5.3 Flash) transcript

A few days ago, rumors started swirling around this mystery model that showed up on Open Router. It was called Ox Alpha, and people were speculating as to what it could be. Maybe it's the new Gemini model. Maybe it's some new continual learning model. Here's the thing, it was fantastic. People were loving it, and it was seemingly comparable to the top frontier models out there. And then we found out what it was. This is the newest open-source open weights model from Z AI.

This is GLM 5.3 Flash. And you know, if it says flash, that means fast, cheap, and smaller than a usual model. And that's what makes it super special. This model performed incredibly well despite its small size. And this might be one of the most exciting model releases to date because you actually might be able to download this model and run it locally, but certainly you're going to be paying a fraction of the price that you're paying for Fable or GPT 5.6 Soul for a model that is insanely capable.

So, I'm going to tell you about the model. I'm going to show you the benchmarks, the pricing, which is crazy. I'll show you how to use it, and some of the demos that I created. But there's something more to this story than just how the model performs. and it's going to have a massive impact on the world. And this video is brought to you by Higsfield. More on them later. So, the first thing you need to know is ZAI is a Chinese AI lab.

And like usual, they're putting out incredible open-source and open weights models. They're telling the world how they're baking the model and they're giving the weights so you can actually use it. You can customize it. You can fine-tune it. You can host it yourself. It is just such an incredible value to the world, these open source models. So this is a relatively small model. It is 320 billion parameters, 18 billion are active.

That's called mixture of experts. It just means it can run really efficiently while also being extremely capable. It outperforms GLM 5.2 which is the last version of the GLM series of models but also the full version of it. And this is a flash version and it is onetenth of the price of the previous GLM 5.2 2 which was already cheap and it is approaching claude opus 4.8 on coding and agentic benchmarks. Now if you were expecting it to be on the same level as a fable it's not.

It's not that far from it but you can't really compare a seven or eight trillion parameter model which is what fable is to a 320 billion parameter model. I mean they're just completely different class sizes. And the fact that it's even close at all is what is so mind-blowing here. So, here are some benchmarks. And keep in mind, they're comparing this new GLM 5.3 Flash to other models that are comparable in size. So, here we go.

Here's Terminal Bench coming in at 84.3. A nice bump over the last full-size 5.2 model. Here's Deepseek. Here's Opus 4.8. Here's GPT 5.6 Terra, which remember Terara is kind of that middle child between Soul and Luna on the GPT 5.6 family. And then here's Gemini 3.7 Flash. Here's Deep Suite, which is probably the most accurate benchmark for how people are actually feeling about a model, how good it is when it's really being used in the real world.

So, here it is at 63.4, a massive jump from 5.2. And we can see quite competitive with 5.6 Terra. Here is Opus. Here's Deepseek. Here's Gemini. Here's GDP val. The OpenAI benchmark testing realworld knowledge work. and it is number one by a large margin. So now let's talk about the benchmarks that I think matter more which is the total cost per task completed and we're going to start adjacent to that with this right here which is a gentic coding performance by effort level.

Now what you're seeing right here is the output tokens per task and then on the y-axis over here we have accuracy. So higher up on the y-axis is better and to the left of the x-axis means less tokens used, more density, more intelligence per token, which is also better. And what we're seeing right here is GLM 5.3. So here's Fable, which is crazy. GLM 5.3 Max sits nearly as high as Fable High and it does use a little bit more tokens, but the fact that it's even getting close and that it is a fraction of the price and I'm going to get to price in a few minutes is kind of just insane to look at.

Now, when you go to max effort on Fable, you do get a massive jump, but you're using a lot more tokens and each one of those tokens is much more expensive. So, here's GLM 5.3 Flash on the artificial analysis intelligence index coming in at 57. Claude Fable 5 is 62. 62 compared to 57. And we are talking about a model that is literally a fraction of the size and price. And I'm able to make videos just like this because of the sponsor of today's video, Higsfield.

Every week there's a new AI image or video model that you might want to use inside of your product. Adding each one usually means a separate API, separate billing, and more code to maintain. Higsfield API solves it. It gives you models like Cling, Nano Banana, VO3, and Higsfield Soul, all with a single API key. You pay per generation rather than paying a monthly fee. So, you're literally only paying for what you use.

And they tell you the cost before the generation. Failed generations are not charged. And as they add new video models, new image models, you don't need to change anything in your code. Higsfield also supports web hooks and concurrent jobs, which makes it easier to run real applications for a team or your customers. Higsfield is one of the fastest growing companies of all time. There are so many people out there who are using it and loving it right now.

And during your first week, you can choose two image models and two video models to get at a discounted price. So go click it down below. Thanks to Higsfield. Now, back to the video. Now, here's where it gets wild. Look at this. So, this is the cost per intelligence index task. This is how much does it cost to actually get things done? And that factors in multiple things. That factors in how many tokens does it need to complete a task?

How expensive are those tokens? So, we see Claude Fable 5 at $3.14. Incredibly expensive. We have GPT 5.6 6 soul coming in at 95, a third of the price of Claude Fable 5. Here's Kimmy K3 at 84 cents, which when we first saw it, that's amazing. Now look at this GLM 5.3 Flash coming in at 9. 9 cents per task completed. And that is just a smidge off of the absolute frontier of Fable 5 intelligence. Not that much off. It's like maybe, you know, 5 7% off of the absolute frontier of intelligence.

And it's about 2 or 3% of the price. That is so crazy to see. Now, this is really the important chart. Intelligence versus cost. And where you want to be is in this quadrant right here. That's why it's green. You want to be as high up and to the left as possible. So, what's really cool is GPT 5.6 Luna Max is incredibly cheap. actually about half the price of GLM 5.3 flash, but for double the price, you get about five points higher on the intelligence index.

So, that's a nice trade-off, but you also get open weights. You also get full control over the model. You can customize the model. So, look, GPT 5.6 Luna is fantastic because it is so cheap, especially after that 80% discount a couple weeks ago. But GLM Flash is a better model. It has a higher intelligence score and it's about 40% more cost. But we're talking about 5 cents versus like 9. So it's really really inexpensive.

So when I say 40% more, both of them are still incredibly cheap. So here's Deepseek V4 Pro coming in at 27 cents. Much more expensive, but also not as good. So GLM Flash is sitting in a great place. If you really care about cost, GPT 5.6 6 Luna still is the number one model in terms of the best trade-off of cost and quality, but GLM 5.3 Flash, man, that offers a lot of benefits. Now, here's something really interesting.

If we look at GPT 5.6 Luna Max, this is the output tokens per intelligence task. Basically, how many tokens does it take to arrive at the same answer as another model? And again, here's Luna at 20,000 tokens on average. Now, if we go all the way to the top, GLM 5.3 Flash is actually one of the most token intensive models out there, which is kind of crazy at 47,000 tokens. So, when you go back to this chart and you see Luna Max here, you see GLM 5.3 Flash here, Luna Max is less expensive because it uses far fewer tokens, less than half the amount of tokens to arrive at the same solution as GLM 5.3 Flash.

So again, these are all the different factors you need to keep in mind when evaluating a given model. But the nice thing is with open weights open source, they will continue to iterate and everybody will have their eyes on it and they will improve it and potentially improve the number of tokens used per task theoretically more quickly than what an open AI can do. Now I'm saying this all in a vacuum. This is all speculation, but I really am a big proponent of open- source open weights for that reason.

Now, a couple last things before I reveal the most shocking part of this entire story. So, number one, it's a million token context. Wonderful. 131,000 max output tokens and max reasoning by default. This is all really good stuff. You know, 1 cent per million cached input. It's it's incredibly cheap, incredibly great. But here's the crazy part. reported by semi analysis. It was serving a 100red trillion tokens per day on purely Chinese chips.

That is crazy. That type of capacity without using a single Nvidia chip is wild. And I've been talking about this on the channel for a little while now. China is developing their own chips. They are co-designing it with the models that are incredible. Plus, of course, the model's small, it's efficient, it's a flash version. So when you pair these things together, the fact that they can have near frontier intelligence at a fraction of the price served entirely on their own infrastructure is quite surprising.

I was not expecting this. So over the past week, we have served GLM53 flash on a large-scale cluster of Chinese AI chips supported by high bandwidth interconnect and a serving stack optimized for the underlying hardware. All of the parts of the AI stack are being co-designed for Chinese models, Chinese hardware, Chinese interconnects. All of these things are being co-designed together and they have capacity now. Hardware efficiency and per token cost comparable to mainstream Nvidia GPUs.

This demonstrates that Chinese chips can support Frontier model inference efficiently and economically at scale. This is all great. This is all really cool news. So, the way that I tested it out, the way that you can test it out is by going to Z.AI, which is being served from China, so keep that in mind. And I've just spun up an API key. I'm not doing anything sensitive. I'm not sending sensitive information. But if you just wanted to try it out, this is probably the most straightforward way.

You can go to Open Router as well. You can go to a bunch of different inference providers because it is an open weights model and so anybody can serve it and they're all going to compete on price. are all going to look for their own optimizations to ek out every penny they possibly can out of that price which of course benefits us. This is the promise of open source and today I'm plugging it into open code. You can plug your API key into anything your project, you can plug it into T3, you can plug it into anything you want as long as it supports an OpenAI compatible API endpoint and the API keys.

And of course, the first thing I wanted to test is a Rubik's Cube simulation. And here it is. It looks hyper realistic. It's smooth. I can grab one of the sides and turn it easily. I asked it to create a bunch of different sliders. We can see right here. You can change all of these. Now, click scramble right there. And it scramles. Everything looks accurate. It's scrambled correctly. The physics are all correct. And then, of course, we can solve it just like this.

Now, all recent models can pretty much create the Rubik's cube simulation, but it's a good baseline to just see how it does. And by the way, we can also change the size of it. We can change the colors. So, these are all settings that it decided to create. So, here's this. I'll do scramble. Now, you can see this is a 5x5 cube. And let's solve it. Yeah, look at that. We can change the turn speed, the scramble length. Let's turn that up.

We can do auto spin. So you can see right there field of view, zoom in and out the glow, the exposure, the key light, here's some shadow, material. We can change the reflection amount, the metaleness, which I don't even know what that is. And then clear coat, so you can make it shinier or less shiny. So very cool. Worked very well. Now, I also wanted to test it against GPT 5.6 Soul. So I gave Soul and GLM 5.3 Flash the same prompts to create a few different demos.

Let me show you those. All right. So, here's the first experiment. Build a beautiful interactive 3D scene featuring seven miniature biome diaramas floating against a deep navy background. So, I'm basically trying to create these little low poly diarama type things. And let me show you the results. All right. So, the final comparison on the left is Soul. On the right is GLM. Now, GLM looks good, but you can just clearly see the winner by far is Soul. just the amount of detail, the coherency of it, it just looks a lot better.

Okay, next I had Soul and GLM53 create five different websites just to see how it does on website design. So, I had it create one about apples, a DJX Spark, Rubber Ducks, Galaxy Zfold, and the Tesla Model Y. Now, here's the website about Apples. It just like kind of plopped a photo of Apples right in the middle. It really does not look good at all. And actually one other thing I want you to keep in mind is that GPT 5.6 soul and within codeex has access to the internet and open code right now kind of doesn't.

So just keep that in mind. But you know the website overall is okay. It's pretty sparse. The colors are okay. But just plopping this image of three apples right in the middle does not look good at all because it overlaps with the text in the background. And here's GLM 5.3. I think this is overall a much better website. Obviously, it does not have access to the internet cuz it did not grab an image of a real apple. But overall, I think it looks better.

Here's something kind of interesting. It actually put an address here. 42 Frost Hill Road in Vermont. Let's see if that's even a real place. Yeah. Okay. So, here it is. It is a real place. I can't tell if it's actually an apple orchard, but yeah, there it is. I wonder why it chose this address. That's kind of wild. All right, so here it is side by side. I think GLM53 won by a pretty big margin on this one. Now, here's the comparison of a website about the DJX Spark.

On the right, we have GLM53. On the left, we have Soul. And once again, I actually think GLM1, it has better design taste than Soul. Soul made up this image of what the DGX looks like. GLM 53 didn't have an image, so it didn't try to make it up. I really like that it has this kind of terminal looking UI here. Uh but yeah, overall Soul is not as good. This is much better. Next, rubber ducks. Here's a website about a rubber duck.

So this is on the left soul on the right GLM. And I actually think they're both really good. This is okay. I don't know what these are. They kind of look like parts of a duck. Uh all of them tend to be very very simplistic. I would still give the overall win to GLM. I mean, the duck actually looks like a duck here. Obviously, Soul was able to pull a real image, but like look at these recreations of ducks on the left, which is Soul, on the right, which is GLM.

Here's a website about a galaxy Zfold. And again, it was actually able to pull up an image. I actually think they both are pretty bad. This is like a very simplistic website. It did pull some information about it, although I don't think these are accurate. This just has completely blank space that doesn't look good at all. Yeah, these are both really bad. And then a website about the Tesla Model Y on the right. GLM. This is embarrassing.

They basically just copied the Tesla website. It looks identical to this. And then somehow created this SVG uh of I don't even know what it is, a limousine. If you scroll down, I mean, the website itself looks good. It's just, you know, more or less a copy of the Tesla website, whereas Soul actually built a website. Now, this one has an image. These two do not. Yeah, they're both quite bad. I don't know if either of them wins.

Okay. Next, I had it put together a PowerPoint presentation about data centers using Forward Futures brand guidelines. And it actually did go to the website forwardfuture.com, downloaded our branding. This is the accurate logo. These are the actual colors. And it's quite good, surprisingly good. Look at this. And this, by the way, is what would show up in the GDP valpoints and real knowledge work. So, here we go. All the colors are accurate.

The text looks good. We see a little M dash right here. Yeah, I mean, this is a great deck. Yeah, this is phenomenal. I'm very impressed with this. And once again, thank you to Higsfield for sponsoring this video. I'm going to drop a link down below so you can go check them out. Click that link, let them know I sent you. Open source is just so important. The fact that we're getting these open-source open weights models out of China is incredible.

And I break it down in full. Go check out that video right