A Year Later, When Kimi and MiniMax Investors Sat Down Together Again

"The craziest possibility may be that humans won't be the only intelligent species on this planet."

By Lili Yu

At last year's 36Kr WAVES 2024 conference, we deliberately arranged a head-to-head between a Kimi investor and a MiniMax investor. At the time, competition among large model companies was fierce. Because both companies' products were more consumer-facing and better suited to dollar-fund tastes, they had raised funding faster and were constantly compared.

But a year later, with DeepSeek's emergence, China's large model landscape has been completely reshuffled. The two are no longer so antagonistic, and their future possibilities have become a new topic of discussion.

In a sense, this is one reason we're reconvening this panel. At the WAVES 2025 conference held on June 11, we re-invited some of the previous guests to participate. They are: ZhenFund managing partner Yusen Dai, Yunqi Capital partner Chen Yu, Gaorong Ventures partner Hu Shuo, and MingShi Capital partner Xia Ling.

At that point, Kimi and MiniMax had been quiet for some time. But in the previous week, both had made new moves: Kimi open-sourced its coding model Kimi-Dev, and its first Agent, kimi-Researcher (deep research), began limited testing. MiniMax, meanwhile, open-sourced its first reasoning model MiniMax-M1 and completed five consecutive days of updates.

These signals all point to what this panel concluded: although all Chinese AI companies have benefited from DeepSeek, the large model war is far from over.

Beyond this, our discussion extended to the hottest investment topics of the moment — Agent and embodied intelligence — including the bubbles within them. We also explored the "bitter lessons" of investing, and the crazy things that will happen in the AI era.

Over the past year, our conversation has evolved from a rivalry between two large model companies to something more like AI versus humanity. In this confrontation, we've also discovered that, in a sense, AI is more like a mirror of humanity. Studying AI is also humanity's ultimate journey of self-understanding.

Below is an edited transcript of the conversation:

Part 01

DeepSeek and the Six Little Tigers

All Chinese AI companies benefit from DS

But the landscape is far from settled

Lili Yu: How do you all view the large model landscape reshaped by DeepSeek, and the possibilities for the Six Little Tigers? And while you're at it, could you share what Kimi and MiniMax are up to? They'd been quiet for quite a while.

Yusen Dai: As early-stage AI investors, we can't perform magic or predict the future. A year ago, we couldn't predict what would happen; the coming year is equally hard to predict.

We invest in a company because we invest in people. We didn't invest in Kimi because of large models, but because of the team. We're in the early days of a technological revolution. They remain one of the AI startups with China's best teams and the most resources.

If we believe AI is something very big, and you have the best team and the most resources, you can still accomplish many interesting things.

When things were especially hot, the Six Little Dragons seemed to have already become dragons; now there's the opposite reaction. I think both are emotional responses. We focus more on whether the team is steadily doing things consistent with their vision and direction. Whether morale has crumbled, whether the team has scattered, whether they're still innovating in the broad direction — these are the real tests. Great companies tend to reveal themselves at moments like these.

From my conversations with them, I've actually found more reasons to be excited. A year ago, it was all about comparing ad spend, comparing users. Now it's back to the technology frontier, back to deep technical insight. I believe this actually suits technical-founder-led teams better. Because comparing ad spend is really an opportunity for the tech giants, so I think they're in pretty good shape.

Chen Yu: AI is an industry that changes by the minute — one year here equals three, five, even ten years in other industries.

Before the Spring Festival, after DeepSeek's release, whether it was the V3 model or R1's reasoning capabilities, the performance was stunning. So it could rapidly capture user mindshare in a short time. In terms of market share for search applications, it's already in a leading position.

But this doesn't mean large models are only about large language models or only about reasoning. Over the past year, we've also seen MiniMax deliver many surprises in the multimodal domain. For example, Hailuo's video generation has very impressive results. More importantly, it achieved commercialization.

Another thing that impressed me was the voice synthesis model. Have you all been on Douyin recently and seen "Daniel Wu teaches you English"? It's actually MiniMax providing the technical support behind it. The first time I heard it, I couldn't tell whether it was a real person or AI. Only later did I learn it was a MiniMax client — I was amazed at how impressive that was.

These details also show that MiniMax has done a lot of interesting exploration over the past year, but not every project has been in the spotlight.

Startup resources are always limited. Even a company like DeepSeek can't cover everything. With limited resources, Yan Junjie chose to bet on video and voice models. On large model architecture, he's particularly focused on linear attention mechanisms, betting on achieving near-infinite context length in the future — this is critical for social companion applications or building Agent systems.

Of course, this means he may have slightly deprioritized reasoning models. Whether this decision is right or wrong, time will tell. But what I want to say is: for a startup, what matters most is how to continuously innovate and break through with limited resources.

Hu Shuo: Objectively speaking, DeepSeek's emergence has indeed made the large model funding environment more bottlenecked. But after DeepSeek appeared, we had several interesting observations.

First, DeepSeek did excellent user education — more users started using AI products.

Second, we found that early this year, whether DeepSeek R1 or Kimi 1.5, both were similar to OpenAI o1 in using reinforcement learning to improve model reasoning capabilities, demonstrating better model performance. This shows that Chinese teams and entrepreneurs have extremely strong research capabilities and technical intuition.

Another finding: we do see model capabilities continuing to emerge, including long-horizon planning and multimodal capabilities. So whether large model vendors or application companies, they can all use better model performance to create better product experiences and user value for users. This is a very good period for acquiring users and generating revenue.

Xia Ling: The fact that the four of us investors can still sit here today shows the large model landscape isn't settled. MiniMax and Kimi are definitely still at the table — otherwise we wouldn't have been invited.

First, looking at the industry macroscopically, DeepSeek has boosted the global standing of China's entire AI industry. All Chinese AI companies have benefited.

A year ago, many overseas LPs questioned whether Chinese model companies could actually build models on par with OpenAI. Today, everyone is confident. Looking further out, if the future global landscape consists of 5-6 large model companies, we believe at least 2-3 will be Chinese large model enterprises. But which ones? That landscape is far from settled.

Take the intelligent electric vehicle industry we've been deeply involved in over the past decade — the landscape has constantly evolved. In 2014-15, nearly a hundred startups were making cars. A couple years ago, just when people thought the first phase of electrification was settled, Xiaomi Auto emerged out of nowhere. And the second phase of intelligence will certainly reshape the auto brand landscape by 2030.

Conversely, DeepSeek is a process of repositioning and reflection for all large model companies. Before DeepSeek v3's release, MiniMax was already adjusting its positioning, making leading model R&D its highest priority. Including ByteDance and other members of the Six Little Tigers, all made positioning adjustments suitable for themselves after R1. As investors, we focus on whether the resources needed to support a company's positioning are in place — funding, people, compute, etc. Another important thing for startups is whether they still have non-consensus views. Like MiniMax's own insistence on Linear Attention from the perspective of Agent development needs.

I believe MiniMax remains a company with ambition, sufficient resources, and its own non-consensus convictions. We have high expectations and are very confident.

Part 02

Signals and Noise in the Agent Era

"Wrapper" has wrapper's value

Still underestimated by a researcher-dominated world

Lili Yu: Many people say the Agent era has arrived. What signals indicate Agents have truly arrived? Compared to tech giants and model companies, where do application companies building Agents actually have opportunities?

Yusen Dai: There's much debate about Agents right now. We do see Agent as a type of AI application that can autonomously plan, call tools, reflect on results, and complete long-horizon tasks.

AI development is a bit like boiling water. Only when water boils can you unlock the steam engine opportunity. With the rapid advances in reasoning, coding, and tool-use capabilities over the past 12 months, the steam engine moment for Agents has arrived.

The two applications we've invested in, Manus and Genspark, I think have a shot at breaking the record for fastest ARR growth in human history.

The value an AI application brings to users operates on several levels:

The first layer is the model layer, call it API if you like. The second layer is context. Context includes three layers.

The first is public context — news, Wikipedia, recently occurred public domain information. The second is organizational context — a company's internal documents, processes.

The third is my conversation history with it as a user, my personal preferences. Public domain is accessible to everyone, but different ways of accessing it often create different value. Finally, what environment you interact with and change — this is built by application companies.

Using a model is like building a car engine. Once the engine is good, context and environment are largely built by application companies. Model companies can't do everything. We'll move from simply needing an engine to needing a complete vehicle.

Of course, many car builders were previously engine builders. Many people say you're just a wrapper, and the model company will eventually come in and wipe everyone out. Many excellent AI applications — people call them wrappers on one hand, and on the other they're becoming ten-billion-dollar companies.

We've discussed internally: where does wrapper value lie, and will foundation model companies enter the Agent market? First, I think foundation model companies like OpenAI will build their own general-purpose Agent products — no doubt. But in specific domains, specific scenarios, when you go deeper, there's corresponding value. If the PPTs you generate are prettier, users will naturally choose you. It's like hiring employees — you hire based on your needs, but employees are diverse; they don't all need to have worked at big companies. So wrapper has wrapper's value — it's just still underestimated by a researcher-dominated world.

Chen Yu: What we've been paying more attention to this year is actually vertical Agents, and the first scenarios to land will likely be in the automotive sector.

For example, ByteDance's Doubao Auto product, built on its large model, can already deploy quickly across over a million vehicles. Users can directly interact with the Agent system in their cars, so people will soon feel in actual applications how Agent delivers value different from traditional productivity tools.

Other domains are also rapidly becoming Agent-ized. For example, the previous wave of SaaS companies is now gradually introducing Agent technology, often pushed by customers.

The biggest difference between Agent and the previous generation of SaaS is this: in the past, SaaS mostly delivered software — the final result depended on how the customer used it; so customers' willingness to pay wasn't always strong. Now, Agent directly delivers results — how well it works is immediately visible. If the results are good enough, customers naturally want to pay.

Why are systems like Manus growing so fast? Because users genuinely feel the efficiency gains. Productivity improvement is visible, results are tangible — so they're naturally willing to pay. This is where Agent's true value lies.

Hu Shuo: I believe AI Agent is a foundation model with reasoning capabilities that can call tools to interact with the real world and autonomously complete tasks. Without tools, it's just a foundation model; without reasoning, it's just a tool.

Most apps and websites we use today are also Agents — they just lack reasoning capabilities. Based on this definition, we see two types of Agents today: the general Agents and vertical Agents that the two investors just mentioned.

From a competitive perspective, I believe in the general Agent domain, tech giants with user scale and leading large model companies have natural advantages. Of course, today we see many excellent startups also exploring and challenging general Agents.

Vertical Agents, because they can accumulate their own unique knowledge base and memory, forming a certain moat, have the opportunity to acquire a certain scale of users based on vertical domains. Even in the next stage, once they have a certain scale of users, they may in turn explore the future entry point for general Agents.

Xia Ling: We began deploying Agent investments in the second half of last year, but the moment we truly felt we must take Agent very seriously, with real urgency, was indeed after DeepSeek exploded. Because DeepSeek is an L2 capability, people began thinking: when will the L3 stage arrive? After communicating with domestic and international large model companies, we felt the Agent L3 milestone might arrive within a year. So very quickly, MingShi internally developed strong urgency about Agent.

In February this year, we gathered our portfolio companies to discuss Agent, actively encouraging everyone to try and think in the Agent direction, including Genspark, Liblib, and Aiyu Intelligence. I think these portfolio companies reacted very quickly and achieved good results. Market consensus formed very fast.

We relatively early proposed a view: in China, building vertical domain Agents shouldn't be treated as tools — they should deliver results. Recently, Sequoia US also proposed focusing on results. Consensus is forming very fast across the industry.

Today, three types of Agents are what we've been consistently focused on.

The first type, let's call it "100x capacity increase" Agents. Whether coding, images, or PPTs — anything that can increase capacity 100x.

The second type is "100x cost reduction" Agents. Compared to the previous type, the results are relatively more clearly defined, and the process workflow is also more deterministic.

This type of Agent may not start as a human-AI co-creation process, but rather tries to let AI complete the entire chain end-to-end as much as possible, ultimately achieving 100x cost reduction. This allows scenarios where ROI previously didn't work to become viable again, changing the entire business model.

The third type of Agent is "completely creating new things," such as AI for science and so on.

These three types of Agents are what we continue to watch; there are opportunities for startups in all of them. But I think the most important thing is that startups must stay focused. We've always emphasized that startups should work in a vertical, professional scenario. Not just focused, but fast iteration — rapid experimentation and iteration on the focused audience or scenario, so they can avoid competing with tech giants.

For entrepreneurs, this may be the opportunity for startups to break through in the Agent era.

Part 03

Embodied Intelligence in Controversy

Hoping there's beer after the foam

Lili Yu: Another investment hotspot this year is embodied intelligence. Early this year, because Allen Zhu exited some projects, it triggered industry discussion about whether it's in a bubble. What do you all think? Which opportunities are near, which are far?

Yusen Dai: The bubble is definitely large. My intuitive feeling is that both investors and entrepreneurs are very enthusiastic. Some people have an attitude of "since investors are giving me money, I'm here anyway, might as well do it" — the bubble is not small.

We talk about the DeepSeek moment. DeepSeek didn't fall from the sky. First the transformer technical direction became relatively clear, then large-scale training with GPT, gradually scaling up. From GPT-1 to GPT-2, GPT-3. Then from technology to product — it wasn't GPT-3 that went viral, it was from GPT-3 to InstructGPT to ChatGPT, it was the product landing that exploded.

Looking at this process, embodied intelligence may still be at the stage of figuring out which technical direction is clear, and only then scaling up, then product landing.

Compared to language models, it's at a very early stage — I think there's definitely a bubble. But we do early-stage investing; inevitably we'll invest in many bubbles, or participate in them.

But there are two types of bubbles in the world. The first believes the past will continue into the future — for example, people who thought real estate would always rise. The second believes the future will be very different — for example, tech bubbles. The most famous is the dot-com bubble.

I think the second type of bubble is a good bubble, because you believe the future will be very different. Whether building internet infrastructure back then, or buying compute now, this bubble has strong positive spillover effects.

People participating in the embodied intelligence bubble — entrepreneurs, investors — should be mentally prepared. We hope there's beer after the foam.

We're very focused on this, very focused on founders who would do this even before it was hot, even without investor money, even paying out of their own pocket. Not founders who wait for investors to set everything up. That's mainly our attitude.

Chen Yu: Embodied intelligence does have quite a bubble right now. If you measure it with traditional PE or PS metrics, it's basically indefensible. But this reminds me of autonomous driving about ten years ago.

The 2016-2017 wave of autonomous driving startups — many companies started doing L4 autonomous driving. At that stage, most were still at the laboratory level, only able to complete basic tasks. Today's embodied intelligence is somewhat similar — high costs, unclear technical paths, but valuations and funding rounds rising very fast.

Autonomous driving later went through roughly a 3-5 year plateau. Funding nearly stalled then, because everyone had poured in a lot of money early on but didn't see the expected technical breakthroughs or commercial progress. During this process, startups began to differentiate: players with sufficient funding and solid technology could stay at the table, waiting for the commercialization inflection point that might come ten years later; while companies without money and lagging technology basically couldn't survive to the end.

When true commercialization arrives, technology will be more mature, costs will come down, and products will truly have the conditions for scaled deployment.

So for us as early-stage investors, if we don't invest now, there won't be many opportunities later. We hope to pick a few promising horses while the industry is still unformed — with both technical capability and funding capability. It doesn't matter if the direction is unclear; we can simultaneously deploy across differentiated routes in hardware, software, etc., and ultimately the true winner will emerge.

Hu Shuo: Regarding bubbles, it may not be a pejorative. We happened to do internal research: in past technological revolutions, because capital markets exist, they accelerate technology's potential release. So for many new technologies, bubbles are an inevitable path.

The only requirement for investors is to enter earlier and exit at a good timing.

Returning to embodied intelligence, we pay more attention to achieving commercialization and scaled deployment. Indeed, today we see that in certain vertical domains, because hardware costs continue to decline and AI capabilities continue to improve, the inflection point for commercialization has appeared. Several industries' market sizes have multiplied several times in the past 1-2 years — this is an industry we're very bullish on this year.

Xia Ling: I think we should first clarify things at the facts level. Today there are several relatively clear facts about embodied intelligence:

First, embodied intelligence so far has not found its scaling law in the pre-training stage — what data to use, based on what algorithms, to achieve scaling law performance.

Second, on the hardware side, there has indeed been great progress, but there's still insufficient engineering maturity. For example, costs — whether humanoid or upper-body humanoid with lower-body chassis — are basically $80,000-100,000, still quite high. Another point: current mean time between failures struggles to meet true commercial deployment requirements.

Third, embodied intelligence today can carefully select some landing applications, but large-scale generalized commercial scenarios are still very difficult.

These are some basic facts. Based on these facts, different institutions have different investment strategies. Some will only invest at PMF; some will invest more aggressively. This relates to each fund's capital characteristics and investment strategy.

I think as long as everyone's understanding at the facts level is correct, it's hard for us to judge whether different institutions' approaches are definitely right or wrong.

We focus more on ourselves. MingShi is an early-stage tech fund. We still hope to invest in the best founders at technology inflection points or commercial inflection points. I believe embodied intelligence is indeed a long-cycle track; technical and commercial inflection points will continue to emerge. We've deployed before, and as inflection points approach or appear, we'll continue to deploy a batch of such companies.

Part 04

Some "Bitter Lessons" in Investing

By end of 2026 at the latest

All promising AI large companies will have emerged

Lili Yu: You've all experienced the mobile internet era. Have you noticed differences in investment paradigms between these two eras? In AI technology progress, we often mention so-called "bitter lessons." In investing, do you have similar summaries?

Yusen Dai: Our fund has had the same logic for over a decade: invest in people. Whether investing in internet, consumer, or AI, this logic persists. As technology changes faster and faster, the invariant you can hold onto is still people.

Our relatively bitter lessons in the past were generally thinking too much about the thing — business models, future direction, studying technical details — while subjectively neglecting people's growth potential.

Now we're actually leaning more toward people. We're indeed not AI experts, but for people, regardless of how technology changes, genuine interest in something, true persistence, true passion — these are key conditions for getting things done.

Lili Yu: At the people level, are there obvious differences between the two eras?

Yusen Dai: I think foundational learning ability, leadership, and innovation are all quite similar.

Judgments about people, influenced by recent project performance, create recency effects. For example, last year everyone was looking for scientists like Yang Zhilin, requiring scientists who understood business. Now people look for someone like Red Xiao, who has always done applications and products.

Many externally different types of entrepreneurs can succeed, but their core is similar — love for technology, for new things, always persisting, also very focused on product and business. These core qualities are shared.

Chen Yu: Every investment institution, every investor, has their own persistent paradigm. Yusen is the typical "invest in people"; we're more "invest in technology."

We've always focused on the fact that AI development highly depends on foundation model capability evolution. As long as underlying models continue advancing rapidly, they will inevitably create new demands for the entire infrastructure, and may even make previously impossible application scenarios feasible.

We pay special attention to changes in underlying technical models. The current pace of technical evolution is very different from before — much faster. We sometimes joke that in the mobile internet era, from iPhone's appearance to the rise of a batch of giant companies, there was roughly a 6-7 year window; but in the AI era, this cycle may be shortened by half.

That is, the true window for early-stage AI investing may be only 3-4 years. By end of 2026 at the latest, all companies with potential to become AI large companies will likely have already emerged.

As for reasons for failure, often it's not technology or direction, but people problems. We've also seen some cases — failed companies aren't actually that bad. The hardest are companies that "just achieved a little success." Once they achieve a little, they easily fall into complacency. Everyone feels they made it through their own hard work, so they start fighting for credit, internal friction, and ultimately bring the company down. That's really a pity.

Hu Shuo: I think this generation will be very different from the previous mobile internet. First, this generation's products don't have the scalable cost amortization of serving users. Second, today we don't have the user dividend and market dividend of that era. Third, user needs are already well-satisfied — this is a major difference.

But much can be reused. From the team perspective, examining capabilities in user growth, product experience, and commercialization.

At the same time, there are more requirements for this generation. For example, more closely tracking model performance and flexibly using models, including having international vision and capabilities from day one.

Speaking of lessons, I'd frame it as a difficult problem for us.

I think the difficult problem we face today is: AI capabilities are so strong, why are there still so few users globally using it. So regardless of how awesome the technology is today, what matters more is enabling more C-end users and ordinary users to use it. For us, this is a problem we've been continuously trying to solve.

For example, we just discussed AI Agent. We're indeed pleasantly surprised to find, from my personal experience, that recently I've indeed seen more AI Agents that can autonomously plan tasks, research and investigate tools, execute tasks, and ultimately deliver results.

Xia Ling: I entered the industry in 2014. What I'm more struck by is the continuous growth and iteration of excellent founders in the previous wave of intelligent electric vehicles.

Even the most excellent founders make mistakes, but how long they take to discover mistakes, how they face their own mistakes, and how they lead organizational growth — I think this is very important. Another point: you must have your own non-consensus, seeing what others don't see and persisting in it. This was very evident in our previous journey accompanying Li Auto's growth.

Seven or eight years ago, during Li Auto's difficult fundraising period, some investors told Li Xiang: why not abandon extended-range, we'll invest if you only do pure electric. But Li Xiang insisted on doing extended-range in the first phase. Everyone recently saw Li Xiang's extremely insightful understanding of AI. This didn't happen last year or this year — at least as early as 2022, he was personally on the front lines deeply learning about this new wave of AI.

So learning ability, growth speed, iteration speed — these are highly shared among excellent entrepreneurs across different industries and periods.

Part 05

The Crazy Things That Will Happen

Humans won't be the only intelligent species on this planet

Lili Yu: What crazy things do you think will happen in the AI era that we haven't thought of?

Yusen Dai: To summarize in one word: the "Lee Sedol moment." AI will exceed the best human performance in many domains. Of course, Go did it before; now we find in coding, in mathematics, in many domains it's matching or even reaching top human levels. We'll increasingly get used to AI doing much better than us. This will bring great changes to industry landscapes.

Chen Yu: In the future, each of us may own a batch of AI "workers" — everyone will have their own Doraemon.

Hu Shuo: Today, we're excited to see some products that can do much better than humans, completely result-oriented, even result-based payment. I think we've already seen it — the real emergence of assistants or AI labor.

Xia Ling: What I find quite certain is that humans won't be the only intelligent species on this planet.

Today society seems more science-heavy and humanities-light than before; many universities are even cutting liberal arts. But I think liberal arts will become increasingly important in the AI era. In the process of humans coexisting with AI, human nature and understanding complex society are humanity's most scarce and valuable things.

The craziest thing in the future may be humans not being the only intelligent species on this planet. But what most shouldn't be neglected is cherishing human nature and humanity's self-exploration.

Lili Yu: This is a very interesting and precious reminder. Studying AI is, in a sense, very much like humanity's ultimate journey of self-understanding. And how fortunate we are to catch up to such an era of great transformation, to participate in and witness all of this together.

Chen Yu: What's truly amazing is that AI is now also learning to "deceive." I initially thought it would be like the Trisolarans in The Three-Body Problem — incapable of deception, without concealment. So AI indeed reflects human nature; in a sense, it's humanity's ultimate tool for understanding itself.

Image source | WAVES 2025 event site

Recommended Reading

The flow of money, the rise and fall of people