Moonshot AI: The “1%” Exception to the Idea That Capital and Compute Decide Everything

Moonshot AI: The “1%” Exception to the Idea That Capital and Compute Decide Everything

2025-12-09

On November 30, ZhenFund hosted a ZhenCampus discussion on AI entrepreneurship at Tsinghua University. ZhenFund Managing Partner Dai Yusen was joined by Zhang Yutong, President of Moonshot AI (Kimi), Zhang Huaiting, Founder of AiEDU, and Zhang Tao, Co-Founder of Manus, for an in-depth conversation on innovation and the future of AI.

On November 6, Moonshot AI released Kimi K2 Thinking, a new generation of open-source thinking agent trained around the principle that the model itself is the agent. It surpassed leading closed models, including GPT-5 and Sonnet 4.5, on benchmarks such as HLE and BrowseComp, setting new state-of-the-art results. Since its beginnings in 2023 as a Tsinghua startup team named after a Pink Floyd album, Moonshot AI has continued to explore the limits of intelligence.

Zhang Yutong leads the company's overall strategy and commercialization. Previously an angel investor in technology companies including REDNote, Infinigence AI, Black Lake Technologies, Liblib, Galaxea, and GravityXR, she earned her undergraduate degree in electronic engineering at Tsinghua.

Starting with the release of Kimi K2, she discussed improvements in model performance over the past year, the agent product experience, and the team's thinking. Even with a valuation only 1% that of other overseas model companies, 1% of their capital investment, and 10% of their staffing, she still believes: "For a long time, the idea that training a model requires a million GPUs and a trillion dollars has been a wall built by the prevailing narrative. Startups need to return to foundational technology and bet on innovation to rewrite the training paradigm."

In her view, "AI is more than an ordinary tool. It amplifies human civilization and is a key to exploring the world of the future." Kimi has brought together people with an appreciation for independent thinking and a curiosity that seeks the truth. They continually test their ideas, taking the genuinely good ones from research to deployment at scale. ZhenFund invested in Kimi's angel round in 2023 and has accompanied the team in exploring technology's dark side of the moon ever since.

The following is the full transcript of her remarks.

I am very happy to join ZhenFund's Tsinghua campus visit today.

From Kimi's founding, we established our mission: to explore the limits of intelligence. Early Kimi users may remember the line on the first version of our website: "Seeking the optimal way to turn energy into intelligence."

Today, I would like to share what we have done in pursuit of that idea.

I imagine many of you here are our users. Has anyone tried our new Kimi K2 Thinking model, launched in November? Its defining feature is that it can think while using tools. I believe this is an important paradigm shift this year: AI products are moving from question-and-answer chat to a more agentic experience——proactively breaking a task into multiple steps, then calling tools step by step to complete it. This is possible because the model itself has gained that capability.

01

Model Capabilities Are Still Improving Rapidly

In overall agent and reasoning performance, Kimi K2 Thinking benchmarks against the world's state-of-the-art models, including GPT-5 and Claude Sonnet 4.5. For example, Humanity’s Last Exam consists of interdisciplinary PhD-level problems that require multistep reasoning and tool use to answer, and Kimi K2 Thinking performs better. It is also highly competitive on other key benchmarks, including BrowseComp, which requires breaking down questions and conducting extensive information retrieval, and the software engineering benchmark SWE-bench Verified.


In July, we open-sourced the trillion-parameter Kimi K2 model for the first time. Its excellent pretraining earned it the description “big and beautiful.” In September, we improved agentic tool use and coding performance. By November, just two months later, further improvements in multistep thinking and agent tool use had produced substantial gains across many benchmarks with K2 Thinking.

This is what excites us most about model development: it is still in a phase of rapid improvement.

We have also received considerable third-party benchmark feedback. On LMArena, a blind-testing arena focused on models' actual performance, Kimi K2 Thinking is the best-performing open-source model. In Stanford Professor Percy Liang's comprehensive HELM evaluation, Kimi K2 achieved the best result among non-thinking models as soon as it launched in July. We were even more pleasantly surprised by K2's high rankings in creative writing and EQ-bench, which measures emotional intelligence. I also use K2 Thinking when I have trouble with writing or replying to messages.

Beyond academic and industry benchmarks, we have seen positive feedback from the developer ecosystem. Kimi K2 is very popular on Hugging Face. Vercel's founder and CEO posted on Twitter, now X, that Kimi outperformed other closed models in their internal agent tests. Social Capital's CEO also shared on the All-In Podcast that their portfolio companies have moved a great deal of work to K2 because of its strong performance and far lower cost than leading closed models.


We also see adoption by leading AI applications worldwide. Perplexity, primarily an AI search application, has tens of millions of active users and 280 million monthly website visits. Besides four closed models, the only model on its current list is the latest Kimi K2 Thinking. This also validates our model's technical capabilities in search. Deep search and deep research are important use cases in which Kimi performs particularly well,
and K2 Thinking supports as many as 200 to 300 rounds of tool calls, thinking while using tools to complete complex research tasks.

02

The “1%” Exception to Capital-and-Compute Determinism

The continued rapid improvement in model performance is tremendously exciting to us as founders. Yet throughout our journey, we have also found that, for a long time, people assumed AI model capabilities had a linear relationship with astronomical capital expenditures: stronger models necessarily required more capital.

From the beginning, Kimi has faced skepticism grounded in this narrative that compute and capital determine everything. A common industry refrain is, "Without a million GPUs, you cannot build a large model. Only the big companies can." People often ask, "How many GPUs do you have? Which models? Have you built your own data center?"

The questions themselves assume a very strong premise. But things became interesting when we gradually realized compute resources were not the only story.


A few days ago, Anna, ZhenFund's Founding Partner, sent me a post she saw on X, formerly Twitter. I found it interesting. It said our company's valuation was less than 1% that of other American frontier-model companies, yet we had also built a frontier model.

With less than 1% of the resources, naturally you work differently.

This has happened many times in the history of technology. Hardware evolves rapidly, and software does too, but software often iterates faster. Their joint evolution can also produce unexpected breakthroughs. In a recent episode of the Dwarkesh Podcast, Ilya Sutskever noted that AlexNet used only two GPUs. Truly breakthrough research does not depend on enormous computing resources before it reaches the stage of engineering at scale.

An interesting point behind this, I think, is: When you believe in a different logic, the entire paradigm of how you work changes with it.

When we shifted from focusing solely on the scale of compute to asking how to maximize the efficiency of converting energy into intelligence, the entire paradigm for training large models changed.


One very important thing we did with Kimi K2 was demonstrate, for the first time, that the second-order optimizer Muon could work on a trillion-parameter model, making training both stable and scalable. This improved token efficiency by at least a factor of two——
Improved token efficiency does not simply mean halving training costs. It means extracting more intelligence from the same data.

The real bottleneck in training today is the data wall, not compute itself. So foundational architecture and algorithm innovations continually improve token efficiency. The Muon optimizer had already been invented, but no one had used it to train a trillion-parameter model, because applying it at true scale introduces extremely complex challenges in training stability.

Second, we place great emphasis on jointly optimizing infrastructure and model training. We call this Day-0 Co-Design. Before training even begins, we consider what deeply integrated design will improve efficiency and how infrastructure and algorithm specialists can work closely on joint optimization. We have done extensive work on foundational infrastructure and algorithm innovation, with every improvement compounding the efficiency of intelligence.

Using just 1% of the resources, a different training paradigm has therefore produced an entirely different result——training a frontier model with a cluster on the order of a thousand GPUs.

We are not saying we already have the world's best model, but that is a very important strategic goal for us going forward. What we can already do today is produce the highest intelligence value per unit of compute.


We are also delighted to see more and more popular AI applications adopting Kimi K2, including AI coding products Cursor and YouWare, general-purpose agent Genspark, and almost all the major internet companies you know. Their recognition of Kimi K2 will help us accelerate model iteration further.

03

Define the Agentic Product Experience Beginning with Pretraining

This year, besides substantially upgrading model capabilities, we put more of last year's “model as product” philosophy into practice, vertically integrating models with the agentic product experience.


We focused on several things:

First, we define the agentic product experience starting at the pretraining stage. We include extensive data from real agent scenarios in pretraining, including agent trajectories——the sequences of tool use and multiround planning. We also establish many internal benchmarks, using custom metrics to measure data quality and our distinctive product sensibility. A great deal of the work therefore starts on Day 0.

Second, once a product launches, we can use real-world user experience as a signal for targeted model optimization and continuous iteration. This genuinely connects model capabilities with user experience. For example, how can a model act autonomously within context without requiring explicit human definitions? How can multiround planning and tool use work together to provide a more complete agent experience?

Third, we dogfood our own model and pass those cost savings on to users (Editor's note: this means using their own API, without paying an intermediary's markup). We devoted these resources to giving people a full experience of deep, long-running agents: completing their tasks more thoroughly through dozens or hundreds of tool calls and searches.

We are also preparing more long-running product plans. In user research, we found a trend resembling Moore's law: tasks are becoming increasingly complex and taking increasingly long. People are gradually realizing agents can do more. The ultimate form might resemble working at a company, where you can work continuously for an entire week after a weekly meeting. Today's agents are still far from that ability to handle long-running tasks.  

But we want agents to call more tools, execute longer tasks, and solve problems that are more complex and economically valuable. This is why we launched Kimi's agent mode, OK Computer. Its name was inspired by Radiohead's classic album of the same name.

After launch, we saw users pushing OK Computer to its limits. The current online version supports up to 50 tool-use steps. We will soon upgrade that to 200 –300 steps, since the new model already supports longer tool-use chains. Some users upload enormous Excel files for agents to perform complex analysis on millions of rows. Others upload many files so the agent has full context for a task, so we are continually increasing the number of files that can be uploaded at once.


OK Computer now supports more than 20 tools, including image and audio generation. Its capabilities extend beyond coding to design, product definition, subsequent development, and deployment as a usable service. Our goal is to make OK Computer everyone's full-stack assistant.

04

Join Us in Exploring the Dark Side of the Moon

What comes next?

Beyond scaling compute and model parameters, many other directions remain open. We can address the data wall, for example, using synthetic data to produce higher-quality data with better distributions at scale.

Agents can now call dozens of tools. In the future, could they learn rapidly and adaptively like humans and generalize to thousands of tools? Strong agent capabilities emerged in models less than a year ago, leaving enormous room for growth.

Beyond these scaling directions, we will continue to emphasize model architecture innovation and firmly bet on next-generation architectures. Our recently released Kimi Linear is just one example. More foundational architecture innovations will follow, and I believe their role will grow increasingly important.

At this year's annual meeting, we posed a searching question to Kimi: The arrival of AGI or ASI could bring a better future, with humans and robots exploring the universe together and human civilization making enormous progress. But it could also threaten humanity.

In that situation, if you were a scientist researching AGI or ASI, would you continue developing it?


Let me share Kimi's answer.

Kimi believes AI is more than an ordinary tool: it is an amplifier that can fundamentally change human civilization, an extension of human cognition, and a key to exploring the future world. It can help us redefine human identity and meaning.

It also recognizes AI's risks. "History shows that some technologies can indeed get out of control and must be treated seriously. But that would not make it abandon AGI or ASI. Even with risks, it would choose to continue, because giving up means giving up the potential of human civilization. We cannot stop exploring out of fear of technological risks. History has shown that every technology comes with risks, but humanity has never stood still because of fear. It also sees AI as a mirror for understanding ourselves and would continue developing it because AI represents civilization's possibilities and is the best tool for exploring the unknown, making us smarter, more insightful, and wiser."

Kimi's answer inspired us. If you share its view, we invite you to join us in exploring the dark side of the moon.


I would also like to share what kind of people we are. AI technology and product interaction are developing rapidly, requiring us to learn continually. The breadth and depth of AI knowledge far exceed any individual. We therefore value the ability to learn from few examples, learning speed, and whether someone continually learns and expands their understanding and capabilities. In our view, this is an advantage of human general intelligence over AI: greater learning efficiency.

This year, almost everyone on our team has become a coder. Beyond algorithm and engineering teams, colleagues in marketing, HR, and finance all have their own AI workflows. Everyone has embraced the improvements in efficiency, scalability, output consistency, and results that AI brings. We hope you will also remain enthusiastic about different AI technologies and tools.

An appreciation for independent thinking is equally important. For example, all large models previously used the Adam optimizer, but our researchers recognized the potential of the second-order optimizer Muon. They began experiments with the Moonlight series, with tens of billions of parameters, and eventually applied a second-order optimizer to trillion-parameter model training. We hope you can come up with original ideas, and plenty of them. Not every idea will be good, but we will test them through extensive experimentation. We want to take research into engineering at scale so that genuinely good ideas can be deployed widely.

That is what I wanted to share today. Thank you for your time.

05

Live Audience Questions and Answers

Q: How do model companies, AI applications, and large-model companies view one another? How do model providers see the boundaries of model capabilities? What value do agents actually deliver?

Zhang Yutong: I think it is a highly dynamic process because model capabilities continue to improve. This differs from internet products. When the internet appeared, its technical capabilities and basic form were largely settled, allowing people to build a great deal of depth, thought, and innovation at the product-interaction layer.

The challenge today is that model capabilities are still changing rapidly and dynamically. Applications and products need forward design: imagining a capability that has not yet appeared but will inevitably emerge along the path of technological evolution, then designing interactions and experiences around that direction.

Without that forward design, model capabilities may suddenly leap forward. A product may originally depend on calling a model step by step, but once the model itself has agentic capabilities, much of that work is compressed directly into the model.

The reverse can also happen: a product innovates the experience, but the model cannot yet keep up. We sometimes see forward-looking product interactions that lead the way, while actual usage feels ordinary. The interaction is not poorly designed; the model has not yet reached the expected capability.

We are all exploring this together. We need a better grasp of model capabilities and good forward design so both sides can move ahead faster together.

Q: What is Kimi's core competitive advantage? As competition among large models intensifies, how do you differentiate yourselves from the giants in technology and the market?

Zhang Yutong : I think the most effective way to compete is not to compete. That does not mean giving up; it means finding your own narrative and what you are genuinely good at.

For us, the first strength is foundational technology innovation. There are still many real problems to solve in large models, and fundamentally they require foundational technology innovation. That is exactly where startups have the greatest advantage.

The second is jointly optimizing the foundational and upper layers. Large companies inevitably face organizational barriers and poor information flow. We have just 300 people, and algorithms, engineering, and product can all sit together. Many new ideas can be proposed and tested on the same day. Rapid iteration, vertical integration, and end-to-end optimization are natural strengths of startup teams.

In our product positioning, Kimi users may notice many things we deliberately do not do. We have not pursued everyday entertainment or multimodal generation. We focus instead on the large-model, logic, and agent layers, and on workflows for productivity and complex tasks such as deep research, presentations, data analysis, and website development. These tasks require long-horizon planning and complex tool use, and create greater economic value.

Within that scope, we want to be the best rather than build something bigger that does everything.