China’s First True Frontier Model?

It’s time to wake up

Source: Author using AI

The time has finally come. Zhipu, one of China’s top AI Labs, has released a model that might set a precedent in AI, as it’s the first Chinese AI model, to my knowledge, to be actually competitive in raw performance with top US models.

And no, this is not another of those Chinese models that offer good intelligence at good cost; I’m talking about raw performance. The model is better than anything Google has ever released for Large Language Models (LLMs), and even beats GPT-5.5, OpenAI’s best model, on several benchmarks.

This is no ordinary release. This is a first. Here’s why.

It’s actually a first

Previous Chinese models, such as DeepSeek v4, showed promise but were clearly behind on most benchmarks.

Don’t get me wrong, they were good and, in many cases, already good enough for specific tasks while requiring several dozen times less investment, but still lagged behind the US frontier on almost every benchmark.

But why is this model different?

David versus Goliath

GLM-5.2, today’s protagonist, is a model by Zhipu AI, or Z.ai, one of the primary Chinese Labs.

On Z.ai’s reported benchmarks, it hits 62.1 on SWE-bench Pro, above GPT-5.5 at 58.6 and GLM-5.1 at 58.4, and 81.0 on Terminal-Bench 2.1, a large jump over GLM-5.1’s 62.0, putting it on the same coding tier as the strongest proprietary models while being an open model (meaning you can download it for free).

The strongest claim is FrontierSWE: GLM-5.2 reportedly scores 74.4%, just behind Claude Opus 4.8 at 75.1%, ahead of GPT-5.5 at 72.6%, and well ahead of Opus 4.7.

The model trails Opus 4.8 by only 1%, edges out GPT-5.5 by 1%, and beats Opus 4.7 by 11%.

Source: Zhipu

Perhaps the biggest loser in all of this is Google, which was shown to be completely beaten across almost every benchmark by a model that is presumably way smaller, was trained on way less data, and therefore, should have been an easy victory for Google. But here we are.

Source: Zhipu

In a particular benchmark for AI-generated web design, the model even outscored Fable, the Anthropic model recently blocked by the USG.

Source: Design Arena

A word of caution. Take these numbers with a pinch of salt because, as I always say, benchmarks never tell the whole story, so it’s the reactions from actual researchers/leaders that really should wake people up to the fact that this is no ordinary release.

Two opinions of AI experts I really trust

But how is this possible?

The secret is that there’s no secret

Architecturally, the model is very similar to everything we’ve seen before. A Transformer, almost identical to Claude or ChatGPT.

The secret is that there’s no secret; we know how they did it because they told us.

Just like any other Chinese model, they use sparse attention, particularly the same DeepSeek uses (appropriately named DeepSeek Sparse Attention), the one big thing that separates US from Chinese models.

But what does that mean?

Most LLMs today work the same way; they take in a sequence of words and output the next. You give the model “What was James Garfield, the US’s 20th President's favorite meal?” and the model answers “Squirrel soup.”

This is a true fact, by the way.

But how does it do this?

Well, it does two things: attention and introspection.

• Attention is a process by which the model will take every single word in the sequence, look back, and ask, ‘Did any previous words have any meaningful information I can attend to?’ (e.g., in “The green cup”, ‘cup’ attends to ‘green’ to gain the attribute of “greenness”).
• Introspection is the ability of the model to “query its own knowledge.” Just like you can use your memory to recover a fact, they can, too.

This process of gaining context from the sequence and adding information from knowledge is done several times, up to hundreds for the largest models, which is the reason why we say these models are divided into ‘layers’. But these layers are just repeating this process sequentially multiple times until the model figures out what the next word should be.

For instance, say we have the sequence “What’s the capital of Germany?” Initially, the model has no idea what you’re asking, so it starts using this attention mechanism; ‘Germany’ attends to ‘capital’, and by doing that, the model already knows you’re talking about Germany’s capital. Then it will attend to ‘What’s,’ and the model now knows you’re asking a question.

Now, it knows your intent and knows the topic you’re asking about.

However, the answer is not in the sequence, so either it knows it, or it doesn’t. Therefore, it needs to tap into its knowledge. This isn’t immediate (neither is attention), so it has to be done across several sequential repetitions.

Taken together, you can see this entire process as the model progressively gathering more contextual and epistemic evidence until it eventually “knows” which word it is: ‘Berlin’.

Both are considerably straining computations that require a lot of compute effort, thereby explaining why AIs need such expensive and powerful chips.

In case you're curious, for short sentences, introspection usually dominates computation. For long sequences, it’s attention that is more compute demanding. As the world moves toward agents (very long text sequences), attention becomes the bottleneck.

However, for today’s topic, we‘re focusing only on the attention part, because the introspection component is literally identical between the US and Chinese models, down to the comma, so there’s not much signal there.

And the key difference is precisely how they “look back”.

Enjoying this piece? Then you’ll love my newsletter, which explains AI in first principles and in words you can understand, for those allergic to hype but hungry for knowledge.

Join today.

[embed]Subscribe | TheWhiteBox by Nacho de Gregorio The newsletter to stay ahead of the curve in AIthewhitebox.beehiiv.com

How much attention is needed?

Say you have the following sequence: “The green, uhm, ehhh, mmm, ah yes, cup, was…”

If attention is about finding relationships between words to build an understanding of the text sequence, does ‘cup’ have to attend to ‘uhm’ or ‘ehhh’, or should it simply focus on ‘green’?

Are those particular words valuable, or could that relationship be ignored and that computation saved?

For the most part, at least for Frontier US models, we don’t ask these questions and simply perform dense attention; each word pays attention to all others, without distinction.

This ensures that all relationships are found, but at the cost of a very large amount of required compute. In our previous example, ‘ehhh’ and ‘uhhh’ are attended to by all succeeding words.

And as you’re probably guessing by now, it’s here where Chinese Labs, not nearly as compute-rich as US ones, apply the biggest cost-saving measure and one of the reasons explaining why Chinese models are cheaper (although not the only one, as we’ll see later): sparsity.

In other words, a ‘sparse’ model actually makes an assumption that not all words are worth attending to. Therefore, these models first make a ‘scan’, selecting the “words that matter” and only then commit to those.

This takes the form of an indexer that selects good candidate words that a particular word can attend to and sets a limit. For example, if the indexer only lets you attend to 20 words and you have 1,000 previous words, every word will only attend to its top 20.

I’m not sure this is a good analogy, since I’m no neuroscientist and I’m not sure whether humans have ‘indexers’. So, from a human perspective, the closest analogy is reading with a highlighter.

Imagine you are reading a 1,000-page legal contract and someone asks you a question about one specific clause. The expensive way to answer is to reread all 1,000 pages carefully, line by line, every time you are asked anything. That is dense attention: every word can attend to every previous one.

But humans rarely work like that. First, we skim. We look at headings, keywords, repeated names, dates, formatting, and rough semantic clues to decide which pages are probably relevant. Only after that do we read those selected pages carefully.

A sparse indexer plays a similar role. Before the model spends real compute comparing the current word against all previous words, the indexer asks: “Which previous chunks are likely to matter?” It then selects a limited set of candidates, and the full attention computation is spent only there.

This allows the compute required to perform attention to remain “constant” across sequence length for every word with only a slight performance reduction, making it an irresistible alternative for Chinese Labs, which, as mentioned, aren’t nearly as compute-rich as US Labs.

With regards to DeepSeek’s models, GLM-5.2 adds an extra compute-saving technique called **IndexShare. While a DeepSeek model applies the indexer at every single model layer, GLM-5.2 applies the indexer only every few layers, avoiding this ‘candidate identification’ process from being executed more times than needed. In long sequences, this reduces compute by almost 3x**.

If you want a deeper dive into this indexer mechanism and the overall functioning of sparse attention (and attention in general), ***I wrote a very detailed description here***.

So, it is over? Has China caught up? Well, it’s complicated, but not for the obvious reason.

No, but it doesn’t matter.

No, China has not caught up. That is not the complicated part here because it’s obviously wrong.

The more complicated part is that it doesn’t matter.

A world of extremes

Even today, I keep hearing delusional takes such as “China is 1 year behind.” I think this model settles this as unfathomably wrong.

But social media narrative is always about extremes to get attention, so either you’re in that camp, or you’re in the “China has caught up” one, which isn’t true either.

The key difference, the real gap, is generalization. Chinese Labs, considerably less compute-rich than US Labs, can’t compete on general capabilities with US models; they have smaller models and much smaller training budgets.

This means they have to cut corners; it’s just what it is. And I continue to hold that as long as the current paradigm holds, in that AI progress is mostly driven by compute (both training and inference), China can’t catch up.

But that doesn’t mean they can’t compete, because oh boy, they can. To do so, they focus on particular domains (mostly coding) to be competitive on specific, high-value tasks.

**A simple example of this problem can be seen below, a benchmark, FutureSim**, that measures how well models can forecast future events that occurred after their training using real news articles and such. The idea is to test whether models can work well with new information, basically.

And in such instances, those that require models to work with truly unknown data (i.e, generalization), the performance gap between closed models and open models is gigantic, which proves that Chinese models are reaching excellent performance on specific domains that Chinese researchers decide to focus on, but have glaring deficiencies overall relative to US top models.

But here’s the thing: it doesn’t matter whether China is catching up in overall capabilities. What matters is whether Chinese models are reaching a capability threshold that allows them to be used, while costing 10–60 times less.

I’ve said it in the past, and I’ll say it again, enterprises, the ones that are meant to pay the bills for all this massive AI bet not to come crushing down, require deep models, not general ones.

They yearn for specialized models and don’t give a dime whether the AI model is good at tasks outside the task at hand.

Generalization is key to pushing the frontier, but not for adoption.

And they’ve hit that threshold in many areas, making them the primary option for cost-effective deployments.

This is terrible news for US interests; the way the US has let China dominate the Pareto frontier is deeply worrying, especially since I firmly believe that commodity tokens will represent at least 80% of the total tokens generated in a few years.

If there’s one thing that should be considered a national security for the US in this entire industry, it is this. And this is only fixed by competing.

The green quadrant of the Artificial Analysis Intelligence Index, which shows the most attractive models on an intelligence-per-cost basis, illustrates this perfectly, revealing an insulting dominance of Chinese models in that front.

Source: Artificial Analysis

And right now, that means 80% of the world’s future tokens will soon be coming from Chinese models if US Labs don’t do something about it.

But what can the US do?

Well, for one, at the risk of stating the obvious, US Labs should put a better fight in the mid-sized range portion. Options like NVIDIA’s Nemotron 3 Ultra are promising, but they still feel like a really weak joint effort relative to what China is doing.

Here, I would want the USG to pressure these Labs to publish more open research and models, especially if the latter group expects liquidity from the Government, which I believe will happen eventually. That liquidity shouldn’t come for free, and they should force these Labs to give back.

If anything, as these Labs go public, they should at least have the decency to give back after they stole all of our data. But maybe that’s just me and how I would behave if I were them.

But the US Government can do something else that goes to the heart of my biggest concerns: capital costs.

The capital gap

People think China is subsidizing token prices. No, it isn’t. And we know this for a fact because those same Chinese models are being served by US inference providers at mostly the same prices.

It’s subsidizing infrastructure.

When evaluating the total cost of ownership (TCO) of a token (generative AI models are charged on the token, which can be a single word in chatbots, a patch of pixels in an image, etc.), 90% or more of the total cost is capital.

For every token that OpenAI or Anthropic generates, only roughly 10% is the cost of serving the token, and 90% is attributable to investment costs.

This changes the picture quite a bit. It’s commonly known that, for every gigawatt of compute, you need to deploy $50 billion in capital costs. But in the recent Computex event, Jensen Huang (NVIDIA’s CEO) projected that the number would grow to $100 billion/GW, doubling.

For this reason, I always insist that rising token prices are mostly due to capital costs; seeing NVIDIA grow revenues 50% quarter-over-quarter while keeping gross margins at an absurd 75% is bad news for everyone but Jensen.

But in China’s case, the picture is completely different.

Based on a report, China’s comparable facility build cost appears to be closer to $6–10B per GW before accelerators, versus roughly $50B per GW for a fully equipped US frontier AI data center. Chinese construction and power security is much, much cheaper, around two times less.

Of course, this is not apples-to-apples, and the missing variable is chip costs. But these are also considerably cheaper.

This implies that every Chinese token incurs three times less capital cost. Knowing it’s the majority of the global cost, no wonder why they are cheaper!

All things considered, this is my primary policy recommendation: find a way to reduce capital costs considerably from their current level so that US Labs can be more competitive in pricing.

Otherwise, panic will soon come.

I share similar thoughts in a more comprehensive and simplified manner on my LinkedIn (don’t worry, no AI-generated content there either). As a reminder, you can also subscribe to my newsletter.