Beyond DeepSeek: Inside China’s New Generation of AI Companies

Five months ago, when Junyang Lin left Alibaba’s Qwen team, one question immediately followed him: what would one of the young researchers who helped shape the Qwen model family do next?

The answer came this August. Lin founded Pragmatik Labs in Shanghai, with an ambition to build “next-generation agents spanning both the digital and physical worlds.” The path is already familiar in Silicon Valley: leave a frontier AI team inside a tech giant, then build an independent company around the next big technical bet. Now, the same pattern is becoming increasingly visible in China.

At almost the same time, another Chinese AI startup was generating a very different kind of attention in Silicon Valley. Moonshot AI’s Kimi K3, released in July, quickly drew interest from developers around the world. The 2.8-trillion-parameter open-weight model performed strongly in coding, agentic tasks and long-horizon work, highlighting a broader shift: Chinese models are increasingly competing for global developers through open weights, lower costs and rapid iteration.

Put these two developments together, and China’s AI story starts to look much bigger than a race to catch up on model performance.

A new generation of technical founders is emerging. They come from quant funds, university labs, overseas research institutions and China’s biggest internet companies. And they are making very different bets. Some are focused on AGI and frontier models. Some are betting on open ecosystems and global developers. Others are starting with multimodal consumer products or enterprise applications.

DeepSeek, Moonshot AI, Zhipu AI, MiniMax and MAAS represent several of these different paths. Their backgrounds, technical strategies and business models vary widely, but together they offer a useful window into where China’s AI industry may be heading next.

DeepSeek: Why Did a Quant Fund Founder Start Chasing AGI?

If one company has done more than any other to change how the outside world thinks about Chinese AI, it is probably DeepSeek.

Its founder, Liang Wenfeng, is also one of the least conventional entrepreneurs in China’s new AI generation.

Liang studied information and communications engineering at Zhejiang University, but later went into quantitative investing. In 2015, he co-founded High-Flyer, a quant fund that used machine learning to identify opportunities in financial markets.

Quant trading is naturally compute-intensive. Long before the current AI boom, High-Flyer was already building GPU clusters and investing in AI research.

So when the large-model era arrived, Liang already had two things most AI founders would love to have: access to serious computing resources and a profitable quant business capable of funding long-term research.

DeepSeek grew out of that foundation.

What makes the company unusual, however, goes beyond models such as DeepSeek-V3 and R1.

From the beginning, Liang appears to have wanted to build a very different kind of organization from the typical Chinese internet company.

He rarely appears in public. He does not spend much time on stage talking about grand commercial visions. And he has shown little urgency to turn DeepSeek into a sprawling “AI super app” with dozens of products.

In a recent multi-hour discussion with investors, Liang gave one of the clearest explanations yet of how he thinks about the company.

His central goal is simple:

DeepSeek wants to increase the probability of reaching AGI.

That idea helps explain many of the company’s choices, including some that can look surprisingly uncommercial.

Liang sees several major steps between today’s language models and genuine general intelligence.

The first is reasoning.

Models need to do more than predict the next token. They need to break down problems, plan and reason. The progress in reinforcement learning and reasoning models over the past few years is part of that transition.

The next step is agents.

AI should be able to move beyond the chat box, use tools, interact with environments and complete tasks.

But Liang does not see agents as the end point.

He is especially interested in continuous learning.

Today’s models are largely frozen after training. They do not learn from the world continuously in the way humans do. A future AI system, in Liang’s view, would need to keep learning from experience and eventually move toward self-improvement, where AI systems help improve the next generation of AI.

The roadmap looks roughly like this:

Reasoning → Agents → Continuous Learning → Self-Improvement

That long-term focus also explains DeepSeek’s unusual degree of restraint.

Video generation may be hot. DeepSeek does not necessarily need to do it.

3D may be hot. It can pass.

A super app may be strategically attractive. It still may not be worth pursuing.

Even with a hugely popular consumer product, Liang does not appear especially interested in winning the title of China’s biggest AI app.

His filter is much narrower: does this move DeepSeek closer to the problems it believes matter for AGI?

That mindset is rare in an industry dominated by fundraising, user growth, revenue targets and valuation pressure. DeepSeek has repeatedly shown a willingness to decide what it will not do.

The same restraint appears in its approach to open models.

Liang remains supportive of keeping DeepSeek’s most advanced models open. In his view, the real capabilities of an AI company extend far beyond model weights: training systems, inference optimization, compute efficiency, engineering and research organization all matter.

If a company’s only moat is that nobody can see its model, that moat may not be very deep.

Liang is also unusually direct about the gap between Chinese and American AI.

He increasingly sees compute as the biggest constraint. Chinese teams can compete at the frontier technically, but American labs still have access to far more GPUs, data-center capacity and capital.

That helps explain DeepSeek’s obsession with efficiency.

When compute is limited, efficiency stops being a nice benchmark result.

It becomes a survival strategy.

And that may be DeepSeek’s biggest impact on the industry so far: it has forced people to reconsider whether frontier AI progress must always depend on ever-larger amounts of capital and compute.

Moonshot AI: How Did a “Star Student” Turn Kimi Into a Silicon Valley Talking Point?

If Liang Wenfeng looks like a quant investor who unexpectedly found his way into frontier AI, Yang Zhilin looks much closer to the archetype of an AI-native founder.

Yang studied at Tsinghua University before completing his PhD at Carnegie Mellon. He later worked at Google Brain and Meta. Among classmates and fellow researchers, he had long carried the kind of reputation that tends to attract words like “brilliant” or “prodigy.”

In 2023, while still in his early thirties, he founded Moonshot AI.

The company’s first breakout product was Kimi.

At a time when many Chinese AI companies were still competing to build something that felt like a local version of ChatGPT, Kimi found a more specific angle: very long context.

It could read research papers, financial reports, contracts and even entire books in one session. For many Chinese knowledge workers, Kimi became one of the first AI tools they actually wanted to use every day.

By 2024, it was one of China’s hottest AI products.

But the more interesting part of the story began after the easy momentum ended.

Kimi’s rapid growth brought outages, product competition and pressure to monetize. Then DeepSeek’s breakout in 2025 raised an even harder question: could an independent startup that still needed to keep raising huge amounts of money for model training stay competitive?

A Financial Times profile of Yang described a fairly aggressive strategic reset. Moonshot reduced its emphasis on short-term commercialization and market expansion, redirected resources toward model training and research, and moved toward a more open model strategy.

By 2026, the results were becoming visible.

Kimi K3 quickly gained attention among global developers after its release. With 2.8 trillion total parameters and strong performance in coding, agents and other complex tasks, the model was competitive enough to trigger serious discussion in Silicon Valley.

More American companies and developers are now experimenting with Chinese open models from DeepSeek, Kimi and Z.ai for a very practical reason: the models are increasingly good enough, and often much cheaper.

That makes Moonshot an interesting test case for a broader question:

Can an independent Chinese AI lab genuinely operate at the global frontier?

Kimi K3 has made that question much harder to dismiss.

Zhipu AI: A Company That Grew Out of a Tsinghua Lab

If Moonshot represents researchers leaving academia to start a company, Zhipu AI followed a somewhat different path:

the lab itself gradually became a company.

Zhipu traces its roots to Tsinghua University’s Knowledge Engineering Lab. In 2019, professors including Tang Jie and Li Juanzi helped commercialize the team’s work, initially around knowledge graphs.

The company then moved early into pretrained large models and eventually built the GLM family.

That origin still shapes Zhipu’s identity.

The company has a distinctly academic feel.

DeepSeek is closely associated with a highly visible founder philosophy. Kimi first became famous through a mass-market consumer product. Zhipu feels more like a research organization that kept expanding outward: GLM, ChatGLM, enterprise models, agents, open models and government and corporate customers.

Tang Jie himself also looks different from the typical technology founder.

He spent much of his career researching knowledge graphs, data mining and artificial intelligence before moving more deeply into business. Today, CEO Zhang Peng is more visible in day-to-day company operations, while Tang is still closely associated with the company’s technical direction and long-term vision.

That model eventually took Zhipu to the public markets.

The company began preparing for a listing in 2025 and went public in Hong Kong in January 2026, becoming one of the first Chinese foundation-model companies to enter the public equity market.

At the same time, Zhipu has continued to expand its enterprise AI business while investing more heavily in open models and compatibility with Chinese AI chips.

The company therefore represents a very Chinese version of a familiar Silicon Valley question:

Can a top university AI lab grow into a major technology company?

Around Stanford, MIT and Carnegie Mellon, that transition has happened many times.

Zhipu may be one of the clearest signs that a similar ecosystem is taking shape in China.

MiniMax: Building Models Is Not Enough — People Have to Want the Products

Yan Junjie’s story is different again.

Before founding MiniMax, he spent years at SenseTime and became one of the company’s youngest vice presidents. Earlier in his career, he had worked on large-scale speech recognition at Baidu.

During that period, he became convinced of a principle that would later reshape the entire AI industry: more data, more compute and larger models could lead to surprisingly predictable improvements in capability.

At the end of 2021, Yan and a group of former SenseTime colleagues founded MiniMax in Shanghai.

The timing was bold. ChatGPT did not even exist yet.

MiniMax also avoided betting everything on text chat.

It moved into multimodality early, eventually covering text, speech, video, music and AI characters. Products such as Talkie and Hailuo AI brought the company into contact with global consumers earlier than many model-focused competitors.

If DeepSeek often feels like a research lab, MiniMax has always looked more like:

model company + product company.

By 2026, that strategy was beginning to show commercial results.

MiniMax listed in Hong Kong in January. Its 2025 revenue grew 159% year over year to $79 million, with more than 70% coming from outside China. Yan has since said that the company wants to remain both a model maker and a product platform.

Those numbers are still small compared with OpenAI.

But they reveal something important:

Chinese AI companies do not necessarily have to rely on the Chinese market.

MiniMax is one of the clearest early tests of whether global consumers are willing to pay for AI products built by a Chinese company.

MAAS: Bringing Large Models Into the Enterprise

DeepSeek, Kimi and MiniMax are closely associated with frontier models or consumer AI. MAAS is pursuing a different opportunity: bringing large-model capabilities directly into enterprise workflows and industrial settings.

MAAS is building an enterprise-focused AI stack covering foundation models, AI infrastructure and industry solutions.

One of its core technologies is a proprietary large language model based on a Mixture-of-Experts, or MoE, architecture. The goal is to balance model capability, inference efficiency and deployment cost by activating different expert networks for different tasks.

For enterprise customers, this matters.

A few extra benchmark points are often far less important than data security, deployment cost, domain knowledge, reliability and the ability to integrate with existing business systems.

That is the gap MAAS is trying to close: moving large models from impressive demos into real production environments.

The company’s direction also fits the background of its CTO, Dr. Zhifeng Li.

Li has a PhD in physics, and his career reflects the mindset of someone trained in the hard sciences: start with mathematical models, computation and underlying technical principles, then move gradually toward engineering and industrial applications.

His career can be understood as a move from theory into practice.

The key question is straightforward:

How do you turn complex technology into systems that can actually run, deploy and create value?

That philosophy is reflected in MAAS’s technical strategy.

The company is going deeper into model architecture, computing infrastructure and enterprise platforms rather than relying only on off-the-shelf models to build lightweight AI applications. The goal is to create an AI stack that can continue to evolve under its own technical control.

Within China’s AI ecosystem, that represents another important path.

Some companies want to build the strongest general model. Others want to own the consumer entry point. MAAS is focused on AI that enterprises can deploy, integrate and keep using over time.

As the industry moves from “whose model is stronger?” toward “who can actually create durable business value?”, enterprise-focused AI companies may find a much larger opening.

Li’s own story fits that transition well: a technically trained physicist moving from theory into industry, and trying to turn AI from a research capability into productive infrastructure.

 

 

China’s AI Race Is Becoming More Diverse

DeepSeek is trying to push toward AGI through better algorithmic and compute efficiency.

Moonshot is using open models to win global developers.

Zhipu is turning university research into a foundation-model business.

MiniMax is betting on both models and global consumer products.

MAAS is focused on getting large models into real enterprise production environments.

They are not following the same playbook, and they will not all necessarily succeed. But the diversity of these strategies is itself a sign that China’s AI ecosystem is becoming more mature.

The United States still has the world’s deepest pools of frontier compute, top research institutions and technology capital. Those advantages will not disappear anytime soon.

China, however, has a different set of strengths that are becoming harder to ignore: a huge engineering workforce, a complete manufacturing and supply-chain base, a massive application market, and a growing number of teams willing to take long-term risks on foundation models.

More importantly, Chinese AI companies are gradually moving from followers to active participants in shaping parts of the global AI market.

DeepSeek has challenged assumptions about the cost of reasoning and the economics of open models. Kimi is gaining attention from developers outside China. MiniMax is testing whether Chinese AI products can win paying consumers overseas.

Their influence is increasingly crossing China’s borders.

The next phase of AI competition will not simply be American companies fighting one another for first place. Nor will it be a one-directional story of Chinese companies trying to catch up.

It is more likely to become a global competition unfolding simultaneously across models, compute, open ecosystems, products and enterprise adoption.

And to understand that competition, it is increasingly necessary to understand China’s fast-growing AI companies — and the new generation of founders and technical leaders building them.