GLM 5.3 Flash: China Shipped a Frontier AI Model Without Nvidia

0:00 Two bills for the same job. A million tokens in a h 100,000 out. One day of an agent reading your code. The bill on the left comes to $7.50. The one on the right comes to 10 cents.

0:12 And the dime isn't a small model cutting corners. On the main intelligence index, it scores 57. The 750 model scores 63. So that's six points apart and 75 times the money. That gap is what this is

0:24 about. And it isn't really a benchmark story at all. It's a story about cost. The cheap one shipped on the 26th of August from a Chinese lab called Z AI. It's GLM5 3 flash. Its weights are

0:37 sitting on hugging face right now under an MID license. Free commercial use allowed. No strings. Then comes the line that moved the company's share price. The company says the launch ran on

0:47 Chinese AI chips. That means no Nvidia hardware anywhere in the serving stack. No lab has claimed that before at this size. It landed in the same two days as a different headline. Nvidia reportedly

0:58 buying hugging face for 13 billion. One company buying the warehouse where open models live. The other shipping a frontier class model into that warehouse without touching its hardware. So that

1:09 is the price. How do you sell intelligence for a dime? Not by being generous. The answer is engineering and it starts with two numbers. 320 billion parameters sit in that file. Of those

1:20 only 18 billion switch on for any given token. Think of a hospital with hundreds of specialists on the payroll where every patient sees nine of them and no more. You still pay rent on the whole

1:31 building, but you only pay for nine people's time per appointment. That's a mixture of experts. And the split is the entire product because your bill tracks the appointment and not the building.

1:41 What you store is a total. What you pay for is the slice that fires, which is exactly where the word compact stops being true for you and me. It describes what fires per token and says nothing

1:51 about what sits on your disc. The download is 331 GB with the weights already compressed. The biggest consumer graphics card you can buy holds 32. You'd need 11 of them just to hold the

2:03 model before you load a word of context. Open weights here means you can audit it, fine-tune it, and rent a machine that runs it, not that it runs on your desk tonight. So far then, enormous to

2:13 hold, cheap to run, but a sparse model isn't new. Deepseek has shipped that trick for years, and its flash line is cheaper than this one. So the ratio on its own can't explain a dime. The rest

2:25 is everything that happens after training. And that's what the launch coverage skipped. It starts with attention, the most expensive habit a language model has. Every time it writes

2:34 a new token, it looks back over everything it has already read and weighs all of it. For a short chat, that costs nothing. Give it a million tokens, and the looking back is the bill. It's

2:44 why long context is where cheap models get expensive. So they split the job in half. Half of it keeps a running summary of what's nearby. The other half fetches only the far away pieces that matter

2:55 using a small index. Then they had to fix the index. At a million tokens, the thing built to save you money starts costing money. So they average four of its key vectors down into one. They

3:05 called it index pool and it exists because the shortcut had grown into the problem. Here's what that buys. Measured against their own flagship, three times less attention compute and a cash 4.4

3:16 times smaller. That cache is the scratch paper the model keeps while it works. And it's what fills the chip first. So that cut is four times as many conversations on the same hardware. And

3:26 the model itself is 45 layers deep where the previous generation was 92. Half the depth, half the active parameters, more capability. Notice what none of those numbers are. None of them is about

3:38 training. They are all serving numbers. What the model costs after it already exists. Which brings us to the silicon it runs on and to the claim the title makes. What the lab wrote is careful and

3:49 it's worth quoting straight. Over the past week, they say they served the model on a large-scale cluster of Chinese AI chips. Look at the verb they used served and at the window they put

3:59 around it. One week. There's nothing in that post about the model being trained on those chips and nothing about what training cost. That's a real gap and it's theirs to fill, not mine to guess

4:09 at. They don't name the vendor either. Not Huawei, not Camberkin, not anybody. A press write up put a 100,000 chips on that cluster, but the lab's own post gives no number at all. So if somebody

4:21 tells you which chip this ran on, ask them where they got it. What they do give you is the engineering. They built a custom inference engine on top of Edgilang with encoding, prompt

4:31 processing, and token generation split into three pools that scale on their own. And their older Gellum 5.3 helped the engineers write the kernels. the model tuning the machine that serves it.

4:41 The result, three times faster end to end than their own first attempt on the same hardware. And then the sentence that got quoted everywhere per token cost comparable to Nvidia's own

4:52 mainstream cards. Comparable, not better. Chinese Silicon didn't win a race. It stopped being a reason you can't compete on price. Which is the point because this company has been on

5:02 the US entity list since January 2025. It couldn't buy the best chips even if it wanted them. That's a lab that had to build this way and the market noticed the stock jumped. So, is it any good or

5:14 is cheap the whole story? Four things and the last one has a countdown on it. On ZI's own coding benchmark, at maximum effort, it lands at 29.0 against Claude Opus 4.8 29.5. But those are the vendors

5:29 numbers on the vendor's harness. Treat them as a claim, not a verdict. That independent number is the one I opened with, 57 on artificial analysis, six points behind the best model anyone

5:39 sells. It takes images and video natively, which the older ZI models did not, and it holds the same million token window opus 5 and JPT 5.6 Soul already have. Then there's the viral number. Ben

5:52 Davis ran the deep coding benchmark against it and got eight out of 10 and 80% went round the internet in a day. 10 tasks is what he actually ran and he said so himself. On the full benchmark,

6:04 ZI reports 63.4. Both numbers are real. They are just not the same measurement. And then there's the part that costs you money if you skip it. That 10-cent price is 50% off. And Open Router puts the end

6:16 date right on the page. The 9th of September, UTC. Watch the discount work. Zi quotes 4 1.5 cents to run an intelligence test and the testers quote 9. So, we have a price with a date on

6:28 it. Today, it's 67 times cheaper than Opus 5 going in and 100 times coming out. 27 and 40 against GPT 5.6 Soul. After the 9th, have them all and it's still 33 times cheaper than Opus 5 on

6:42 input. But watch what happens to the cheap Chinese model story. Today, it's twice as cheap as Alibaba's flash model. At the list price, the two of them land on the same 15 cents with the Alibaba

6:53 model 3 cents cheaper coming out of everything here. Only Deepseek's flash line beats it on price and that one is text only. Deepseek does have a vision model now. It shipped 5 days before this

7:05 one and it lands at 22 cents in. That's three times the flash price today and half again more even after the 9th. Images are where the cheap option stops being cheap. Here's where I land. If

7:16 you're running agents at volume, long context, lots of tool calls, work you'd be embarrassed to pay opus rates for, this is your new default, and it isn't close. That's 75 to1 on an ordinary day.

7:29 To justify the expensive model, you'd have to believe it's dramatically better and six points isn't dramatic. The 5% it's wrong for anything a person is sitting and waiting on because 50 tokens

7:39 a second is cheap, not fast, and cheap tokens don't help when someone's watching a cursor blink. and I'd still take it at the list price. Twice a dime is 20 cents. The idea this kills is that

7:51 the price of intelligence is set by what it costs to train. It isn't. It's set by what it costs to serve, how many parameters fire, how much memory it holds while it works, and what the chips

8:01 underneath cost per hour. Zai just won all three at once on hardware it isn't allowed to buy from Nvidia. But serving is only half the sentence, and it's the half they documented. So before that

8:12 gets called a decoupling, when a Chinese lab finally says it trained a Frontier model on domestic chips too, what is Nvidia's 13 billion warehouse actually worth then?