Four Choices I Learned from Zhang Peng as Zhipu AI Gains Momentum

Zhipu AI has been on fire lately.

Earlier this year, when Zhipu AI and MiniMax went public around the same time, many people naturally placed them in the same frame: both are major Chinese large model companies, both received new valuations in the capital markets, and both had to answer questions about technical capability, commercialization, and ecosystem positioning.

But recently, the market’s discussion around Zhipu AI has clearly shifted.

It’s not just that the label “first AI large model stock” has been revived, or that its market cap has been re-priced. More importantly, the GLM model line has started to re-enter the视野 of developers and overseas KOLs.

From GLM-4.5 to GLM-5.2, Zhipu AI’s trajectory is increasingly clear: reasoning, coding, agentic abilities, long context, and open weights are all moving towards an agent-native model direction.

Recently, GLM-5.2 has pushed this trend to the forefront. Overseas developers, model leaderboards, and the AI infrastructure ecosystem have started sharing and discussing it. As a result, Zhipu AI has been pulled out of the old impression of “stable, technical, To B” and into a much tougher validation arena.

Screenshot: Z.ai officially released GLM-5.2, putting coding, agentic tasks, 1M context, and MIT open weights on the same line.

So let’s be clear: Zhipu AI is in the spotlight, and the model is genuinely good.

But I don’t want this article to be an investment thesis or a postmortem on “why Zhipu AI became popular.”

Stock prices change daily. Leaderboards change.

What I care about more is another question:

When a company finally gets noticed, the truly worthwhile question isn’t why it suddenly became popular, but what choices it made before that.

This is exactly what I felt most strongly after watching the recent video where Zhang Xiaojun interviewed Zhang Peng.

Initially, I expected a standard narrative from a model company CEO: technology, IPO, open source, commercialization, international competition.

But after watching, my stronger feeling is that this wasn’t a conversation about “how Zhipu AI builds large models.” It was a conversation about “how people should make choices.”

What makes Zhang Peng interesting isn’t that he said many quotable lines.

Quite the opposite. His expression is plain, steady, even a bit slow to warm up. He doesn’t seem to be selling a grand narrative; instead, he’s explaining why a company is willing to take a path that isn’t easily visible.

That’s exactly why, now that Zhipu AI is in the spotlight, I think it’s worth revisiting Zhang Peng’s perspective.

1. Choosing a Path That Isn’t Immediately Visible

Looking back, Zhipu AI’s path has never been the most eye-catching.

It’s not the best at creating C-end blockbuster products, not the best at crafting a founder persona, and not the best at generating emotional value on social media.

For a long time, many people’s impression of Zhipu AI probably stayed at a few words: Tsinghua lineage, tech-driven, model foundation, open source, To B, stable – but not exciting enough.

In the AI world, this is a disadvantage.

Because the industry heavily rewards “momentary impact”: a viral demo, an extreme benchmark, a founder’s quote that gets endlessly circulated, a product experience that suddenly trends.

Zhipu AI feels more like another type of company.

It invests long-term in things that don’t transmit well: model training, engineering optimization, enterprise delivery, open-source ecosystem, cost control, customer needs.

These things are hard to get excited about in one day.

But many important life choices are like this.

Not every right choice gives you immediate feedback. Often, you’re not choosing a direction that will be praised instantly, but a direction you’re willing to endure long-term misunderstanding.

I think Zhang Peng’s mindset might be right here.

If you choose to build the foundation, to do engineering, to layer complex systems, you’re likely to go through a long period where “others think you’re not bright enough.”

This isn’t a strategic failure; it’s the cost of this path.

The first thing ordinary people can learn is:

Don’t mistake “temporarily unseen” for “this path has no value.”

For many career choices, entrepreneurial decisions, and long-term capability building, the hardest part isn’t starting, it’s continuing when there’s no external feedback.

2. Choosing to Turn Research into Products, Not Stop at Papers and Demos

In the interview, Zhang Peng repeatedly returns to a very concrete question:

How do you turn research results into products?

How does a lab prototype become a deliverable system?

Why do customers pay?

How do you account for costs and benefits?

These questions don’t sound sexy, but they are precisely the key character of Zhipu AI as a company.

Many AI companies tend to fall into one of two extremes.

One type only talks about the technical route: parameters, architecture, benchmarks, papers, open source.

The other type only talks about the product story: user growth, viral entry points, emotional value, communication buzz.

When Zhang Peng talks about Zhipu AI, it feels more like finding a narrow path in the middle: the technology must be hard enough, but it must ultimately lead to a real system; the model capability must be leading, but it must ultimately translate into a product that customers are willing to use long-term.

This is also harsh when applied to life choices.

A person can’t stay forever in “I have potential.”

A company can’t stay forever in “My technology is great.”

Eventually, you have to answer the same question: Can your ability be used by the world? Can it be verified by real scenarios? Can it go from a good idea to something others are willing to rely on continuously?

I think this is also the strongest sense of engineering in Zhang Peng.

It’s not about making the ideal bigger, but about constantly translating the ideal into something that can be delivered, can withstand pressure, and can be accounted for.

The second thing ordinary people can learn is:

Don’t stay forever in “I have the ability.” Turn that ability into something others can use as quickly as possible.

If you can code, make a product that can be used.

If you understand an industry, consolidate it into a reusable method.

If you have judgment, let it become a verifiable decision, not just an opinion that stays in conversation.

3. Choosing to Let External Validation Speak for You

Recently, the GLM line is particularly hot, and the most valuable part is not “getting on another leaderboard,” but that it’s starting to be validated by a more external crowd.

Rauch’s tweet is typical because it’s not a vague compliment about a Chinese model, but directly falls on the coding experience.

“Genuinely impressed, almost shocked.”

The Arena chart is even harder: In Code Arena: Frontend, GLM-5.2 (Max) ranks #2.

“GLM-5.2 is the best open model.”

This kind of feedback is important because it’s not internal hype within the Chinese market.

It comes from developers, model evaluation communities, frontend/coding agent users, and actual integrators in the AI infrastructure ecosystem.

When people domestically look at Zhipu AI, they often carry the company narrative: Tsinghua, IPO, first large model stock, To B, open source.

When overseas developers look at GLM-5.2, their first reaction is simpler:

Can it code?

Can it run agents?

Is it cheap?

Is it open?

Can it enter my toolchain?

That’s the value of external validation.

It doesn’t care how many stories you’ve told in the past; it only cares whether you can solve problems now.

The same applies in life. You can explain yourself many times, but what truly changes others’ perceptions is often not explanations, but a continuous stream of external results.

This is probably also why Zhipu AI has been re-priced recently.

It’s not that it suddenly became good at storytelling; it’s that the model’s capability started speaking for it.

The third thing ordinary people can learn is:

Instead of repeatedly explaining yourself, go create an external result that others cannot ignore.

Often, what truly changes others’ perceptions is not your self-introduction, but the continuous appearance of your work, clients, data, reputation, and delivery record.

4. Choosing to Return to Long-Term Questions Even When in the Spotlight

Zhipu AI is in the spotlight now.

The capital market is watching it, the developer community is watching it, overseas KOLs are sharing it, and the secondary market is giving it new imagination space.

But it’s precisely at times like this that it’s easy to simplify the issues.

As if once the market cap goes up, the model makes the leaderboard, and overseas people retweet it, everything naturally falls into place.

I think Zhang Peng’s interview reminds us of something else:

When you’re in the spotlight, the truly difficult problems are just beginning.

First, can technical trust turn into product trust?

Developers thinking the model is good is the first step. But can it consistently and stably enter toolchains, agent products, and enterprise workflows? That’s another matter.

Second, can open-source influence turn into commercial results?

Open source can bring attention, trial usage, and ecosystem diffusion. But what enterprises ultimately buy is not “the existence of weights,” but deployment, inference costs, service capability, scenario adaptation, and long-term reliability.

Third, after going public, can Zhipu AI maintain patience in technical investment?

The public market will continuously look at revenue, profit, cash burn, and growth expectations. The hardest part for a model company is to simultaneously pursue the frontier and translate each capability improvement into clearer business outcomes.

These problems are not easy.

But that’s also the meaning of choice.

Choice is not saying “I’m long-term oriented” when the wind is at your back. It’s being able to return to long-term questions even after you’ve been noticed.

What exactly do I want to achieve?

What am I willing to bear a long-term cost for?

Am I doing this to be seen, or because this thing is worth doing in itself, and that’s why I waited until it was seen?

The fourth thing ordinary people can learn is:

When the wind is at your back, the most important question is not “Can I get even more popular?” but “What exactly do I want to accomplish?”

Because hype can amplify a person, but it can also amplify their wavering.

The more things are going well, the more you need to return to long-term questions.

What I Learned from Zhang Peng

So, I don’t want this article to be “Zhipu AI finally got popular.”

I want it to be another sentence:

After the spotlight fades, what’s most worth seeing is not the excitement, but the choices a person and a company made before.

From Zhang Peng’s interview, I see four choices.

Choose a path that isn’t immediately visible.

Choose to turn research into products, not stop at papers and demos.

Choose to let external validation speak for you.

Choose to return to long-term questions even when in the spotlight.

These four things, when applied to a company, are Zhipu AI’s path.

When applied to an individual, they look a lot like life choices.

Often, what really matters is not whether you stand in the wind, but whether you were willing to do things that weren’t easily visible before the wind came.

It’s also not whether you can tell a beautiful story, but whether your ability can ultimately enter the real world and be repeatedly validated by real needs.

Zhipu AI’s current heat certainly doesn’t automatically prove it has won.

But it at least shows one thing:

Some slow things, as long as they can truly bear weight, will eventually be seen again.

Materials are here for your own verification:

https://www.youtube.com/watch?v=toy8RLeFZ08

https://www.cnfin.com/yw-lb/detail/20260109/4363163_1.html

https://www.stcn.com/article/detail/3971526.html

https://z.ai/blog/glm-4.5

https://z.ai/blog/glm-5.2

https://x.com/rauchg/status/2068517095818809770

https://x.com/arena/status/2066957802741043641