SHF2026｜Mars Xue

Mars (Zizhao) Xue 

Vice President of MiniMax

Thank you for having us. Today it's very honored to be here to share our perspective of this industry, the industry dynamics，the landscape, the evolution and our company's situation in this fast growing industry. Our company's name is Minimax and our slogan is “Intelligence with Everyone” (每个人的智能). Our goal is to share the intelligence with global users, with everyone. So before diving into the details, I would like to share this picture I just generated a to share my feeling as an industry participant in large language model sector. 

Three months ago, I guess many of you guys have experienced the agents product like OpenClaw. It's the first time that people can use large language model to resolve tasks, not simple Q&A. It's like an earthquake. What happens after the earthquake is a category-10 tsunami. Now we are in a tipping point that we start to see the large model can solve real tasks with high economic values. This tipping point, this change is like a category-10 tsunami. And what is strange is look at the people on the shore. Many of the people have not yet realized the impact and the speed of this transition. This is basically what we are feeling right now. We are seeing this big shift, big transition driven by foundation model. And it's heading right toward us, and it's moving very fast. That's why we are very very excited to be in the sector.

This is our company's very interesting history, our origination. We are founded 4 years ago, early in 2022, as the earliest established company in Asia, focusing on financial model research. If you recall the time at the early 2022, actually there was no OpenAI's ChatGPT. It was 1 year before the launch of OpenAI's Chat GPT. 

At that time, our founder’s nickname was IO, which basically means input and output. IO foresaw the tipping point where the general use of models would hit this inflection point. What's the difference? The first sentence in Chinese is next generation AI and its difference. The key difference is that when we talk about the last generation of AI, it was task-specific AI — meaning when you had a new customer or a new task, you needed a new model. But now, AI is generally used, meaning you have one same model to serve every customer, every enterprise, and every user. This is a fundamental change. So let's think about this: in human industry, there are actually very few products that can remain exactly the same and serve global users. Another example, I guess, is Coca-Cola. You have the same product serving the globe, but then you need to worry about the distribution of that product. For the general-purpose model, we have the same model and its general use. It can be used in productivity, and it can be used in creativity. It can handle over 200 different languages, and it's the same model. So it's very scalable. That's the magic of the general-purpose model. That's the first point.

The second point is our very unique definition of general artificial intelligence, or AGI. From day one, four years ago, we defined AGI as an agent capable of passing generalized Turing Tests. This means that, just like human interaction, it is not only about language models — it involves all modalities. So if you break down all the content that humans interact with, it includes language, it also includes visual content such as images and videos, and it also includes audio. So from day one, we have been creating three-modality foundation models: the language model, the video model, and the audio model. And another perspective is, it can be divided into intelligence and creativity or imagination. So these are, in summary, the components of human intelligence. That is our definition of AGI. 

And also we foresaw this is a gradual development process. We have different levels of development from L1 to L5. Let's quickly have an overview of the development of the industry over the past few years after the launch of ChatGPT. Three years ago, ChatGPT was launched. It was the first product that reached the L1 level, which means you could have casual chats with a language model and do simple Q&A. Then, in 2025, we entered L2, which means the models started to have reasoning capability and a thinking process. If you ask a question, the model will say, Let me think about it. What is the chain of thought? What is the rationale behind this? That was L2, which happened last year. Now we are standing at a moment where model capability is quickly approaching L3 in many different verticals. The definition of L3 is agents, which means the model can make a plan before doing a task and can deliver the final results just like a human. Going forward, we see L4 and L5. L4 means the model can have innovations. L5 means it can act as a group CEO to organize very complex structures or organizations.

We are quickly seeing the model improving in, say, coding — getting very close to L3. And we see that in many other knowledge-related areas, the model will also replicate the fast curve, just like in coding, to reach L3 and then go beyond L3 to L4 and even L5. So that is the current evolvement of model capability.

I would like to share two very unique features for this industry. The first one is that in this industry, AGI is all about intelligence. Intelligence—the model capability—is the only driver of this industry. We have a very clear flywheel driven by this model intelligence, and this flywheel effect is very different from all other industries, including internet industries. I guess many of you guys are very familiar with internet industries. In internet industries, the flywheel is basically data and user. You have more user data, so you can have a more accurate recommendation system. Then you have higher customer retention, and then more users — a positive data, user, or traffic flywheel. That is an example of what happens in the internet industry. 

But in the foundation model industry, traffic is no longer that relevant. The key driver is the model intelligence. What this model intelligence means is that every three to six months, we will see the model's capability improve to a higher level. It means the frontier of intelligence is expanded, and some tasks that were previously impossible for the model to solve become possible.

There are several examples. Many of you may have tried image generation models. The latest ones are actually GPT Image 2, which just launched two days ago, and Google's Nano Banana. One year ago, before the launch of Google's Nano Banana, the popular image model was called MidJourney. Before the launch of Nano Banana, if you asked people what makes a better image generation model, people would tell you it must have higher resolution, better aesthetics, and so on. But only after the launch of Nano Banana did people suddenly realize that this image model can understand pixels — it can understand, Oh, this is a screen. It can even understand the text and languages embedded into those pixels. And then people found that image generation is not only for professional users — it is for everyone. People are now using image generation models to even generate PowerPoint slides with language embedded into the images. That is an intelligence jump. Every time you have an intelligence jump, it is creating incremental market — not an existing market, not a zero-sum game. 

Another example is AI coding. Now we know — if you have tried an agent or OpenClaw — that coding is basically for everyone. You don't even need to have learned a coding language. You just use natural language to use the model. But 12 months ago, before these latest improvements in the model, if you asked people, What is your view? What is your estimation of the market for coding? people would tell you it's a 30 million programmer market. What percentage of these programmers will be replaced by a coding model? And then, what is your annual salary? That was basically the market potential for AI coding. But if you look at it from today's perspective, it was all wrong. First, it's not only about programmers — it's for everyone. And next, it's not about replacing some existing market. It is always about creating new productivity. So that is always happening in this industry.

So we are still seeing very huge potential from the scaling law, meaning that if you have larger models, higher amounts of data, more training, and more compute, you can achieve a higher standard of model intelligence. And then this newer version of model intelligence will create a very very huge incremental market. That is still working for the foreseeable future.

The second very interesting feature of this industry is that, given its fast growth rate, if it were any other industry, we would normally see more and more new entrants coming in, but this industry is the opposite. The number of players who can launch frontier models is not becoming larger; in fact, the numbers continue to shrink. Even many big companies — the tech giants — after investing huge amounts of money into this sector, still do not have frontier models. This tells us that the real entry barrier for this industry is not simply investing a huge amount of cash or compute. For example, there are actually many US tech giants and some Chinese tech giants that have spent a lot historically, yet they still have not developed frontier models.

So the real entry barrier for this industry, we think, is actually the iteration speed — or the capability to continuously deliver innovations. Because the industry itself is growing too fast. You need to keep up the pace. Your R&D organization needs to keep up with the pace of this industry to continue delivering frontier models. And now the number of players has already shrunk to a handful globally, mostly from the US and China. So that is the second very unique feature of this industry.

Going forward, we don't think this industry will consolidate into only one or two players, like the searching industry. We don't think that will happen. Why? Because just like different people have different personalities, all the frontier models are now very differentiated — meaning they have different focus areas and different territories where they are strong. If you are a deep user of different models, you can clearly tell the differences and their differentiated strengths. So we do not anticipate that there will be more new entrants. We also do not anticipate that it will concentrate into only one or two players. It will likely be a high single-digit number of players leading frontier model development, and they will have different strong territories and different differentiations. That is our view of the landscape of this industry.

This picture is about the relative catching-up historical evolution for language models, and it's very interesting. First, on the left-hand side is a third-party intelligence benchmark for several frontier models. The fact is that the most leading and most frontier models are currently OpenAI and Anthropic. But what has happened over the past one to two years is that Chinese models — including us and also DeepSeek — have been quickly catching up in model capability. If we look at the vertical gap, where the vertical axis represents intelligence level, the gap is clearly narrowing, and we have a steeper slope. That is for overall performance. If we look at some specific areas where we are focusing, such as coding and agentic capabilities, the gap is even smaller. The bottom line — this pink line — is our model. So over the past half year, we have seen the gap narrowing quickly.

Different companies in this industry actually have different strategies. Most Chinese-originated models have adopted an open-source strategy, meaning the model is transparent — all the parameters, all the weights, and all the algorithms are open and published on platforms like GitHub or other third-party channels. Everyone can download the model for free. In contrast, most of the leading labs from the US use a closed-source method — basically a black box. So that is why we are seeing a clear trend: many regions around the world that want to develop their own foundation models are actually more likely to collaborate with open-source models or Chinese models, because they are more transparent and open. That has been the development over the past one to two years. And that is why we, as a Chinese model, are very excited to continue closing this gap, and we will do so even faster.

Here are some interesting stories from two to three months ago. Everyone now knows OpenClaw. The founder of OpenClaw is named Peter Steinberg. He was a one-person company before he recently joined OpenAI, and he was actively promoting our model on his Twitter. After he built OpenClaw, this project gained attention for several reasons. The first time we met Peter himself was last month during GTC in San Francisco. Starting in early January, he began promoting our model on his Twitter. The first reason is that he was actually building the OpenClaw project using our model — at that time called the M2.1 version. But more importantly, we think the main reason he loves us very much is that our model plays a very constructive role in the penetration of this agent product. Why? If you have tried using OpenClaw by plugging in the most expensive models, after only two to three days, you will receive a $500 to $1,000 bill — after using it for just two to three days. Obviously, not many people can afford that. So our model is actually creating an incremental market, making agentic products like OpenClaw accessible to everyone and usable by global users. I guess that is the main reason why the entire agentic ecosystem loves our model very much.

Another example is that we are seeing very strong momentum from global MNCs and global enterprises quickly adopting Chinese models. Here is an example from Notion, a note-taking application, which has started using our model as their third model — after OpenAI and Anthropic — to be integrated into their agent function.

To quickly wrap up, we have these three foundation models as our base, as our fundamentals. For foundation model-driven products, the product layer is actually quite thin. The majority of the product performance is driven by the model performance. We think that for us — and actually for everyone in this industry — the real product is the model itself. So we are seeing this industry iterating faster and faster. And we have an even stronger feeling about this now. If the model can surpass L3 and reach L4 — meaning it has the capability for innovation — the first thing this model will do is participate in the R&D process of the next generation of the model. It means that, just like human researchers and human engineers, the model can have new ideas in R&D.

Now all the frontier labs, including us, are heavily engaging models in the process of iterating the next version of the model. The positive feedback loop is that if you have a more intelligent version of the model, you can achieve higher productivity to create the next version of the model. This will then start to have a self-recursive effect.

That is why, when you surpass a specific threshold, we start to see acceleration in model development. In the coding sector, it is very clear that we already surpassed that self-recursive threshold — basically at the end of last year. That is why, starting from this year, we are seeing acceleration, specifically in the coding industry. The first downstream application or downstream vertical to have this self-recursive effect is the model development itself. And that is very exciting.

And then, going forward in the next one to two years, we foresee a similar self-recursive effect happening in other areas — say, the workspace. For example, in creating a PowerPoint, or in many of the fundamental skills required for knowledge workers. In these kinds of verticals, as long as the context happens online, we think the model can pick it up and quickly reach top-notch professional capabilities. That is the fundamental reason why this industry is accelerating — because there is a self-recursive engine behind it.

Here is a quick overview to give you a sense of our company. We operate globally. Our users come from over 200 countries, and the majority of our revenue comes from international markets. Over the past two months, we have seen six times growth in model usage. From day one, we have had this four-lineup — for video, for audio, for language as a whole — to provide general artificial intelligence. Our mission is not to become a super-intelligent machine or system. What we want — our mission — is actually to share the benefits of this industry and of our company's development with everyone. That is why, from day one, our mission has been Intelligence with Everyone. Thank you all. That is all for my sharing today. Thank you.

(This article is edited based on the recording and has not been reviewed by the speaker.)